<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Simone Basso on Medium]]></title>
        <description><![CDATA[Stories by Simone Basso on Medium]]></description>
        <link>https://medium.com/@smnbss?source=rss-5de037daf82f------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*k4GbjtaSVJscexelvEW-pQ.jpeg</url>
            <title>Stories by Simone Basso on Medium</title>
            <link>https://medium.com/@smnbss?source=rss-5de037daf82f------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 03:01:22 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@smnbss/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com" rel="hub"/>
        <item>
            <title><![CDATA[Context Window — October 2026]]></title>
            <link>https://medium.com/@smnbss/context-window-october-2026-f67f719c807e?source=rss-5de037daf82f------2</link>
            <guid isPermaLink="false">https://medium.com/p/f67f719c807e</guid>
            <category><![CDATA[ai-sharing]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Simone Basso]]></dc:creator>
            <pubDate>Wed, 21 Oct 2026 06:41:27 GMT</pubDate>
            <atom:updated>2026-10-21T06:45:58.543Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>The agent stopped needing your laptop, and then it stopped needing you to start it.</strong> DigitalOcean now runs Claude Code and Codex as cloud sessions that pause when idle and wake in about 300ms on any device — his claims, with no pricing or product page attached (<a href="https://x.com/VadimStrizheus/status/2102449430771417589">VadimStrizheus</a>). Danny Postma’s Hetzner-plus-Orca rig over Tailscale drew more bookmarks than likes, which means people are copying it, not applauding it (<a href="https://x.com/dannypostma/status/2106215980968136958">dannypostma</a>). T3 Code now fires agent tasks from a webhook (<a href="https://x.com/jullerino/status/2107510882435789018">jullerino</a>), and Anthropic published the pattern for Managed Agents that read Slack and GitHub on a schedule (<a href="https://x.com/ClaudeDevs/status/2108320296323477683">ClaudeDevs</a>). Last month the software factory became a category. This month its front door got cheap: an event, not a person at a keyboard. I run this exact shape with T3 Code on Tailscale, and our own morning digests still hang off one laptop’s launchd. That isn’t a platform. It’s a single point of failure with a cron entry.</p><p><strong>The agent found its room, and it’s a Slack channel with a standing order in it.</strong> Boris Cherny says Tag, Claude’s Slack integration, writes more than half of his PRs every day and about 100% of his data analysis (<a href="https://x.com/bcherny/status/2103538666597691552">bcherny</a>). The grammar matters more than the percentage. His examples are standing orders: from now on, reproduce every bug posted here end to end, open a fix PR, tag the right team. Last month I said that once a team can see the agent work, review stops being optional. This month the room won. But Tag now also uses your personal connectors — an agent acting with everything one person can read, inside a room everyone can read. Decide whose permissions it runs with before somebody types “from now on” in the wrong channel.</p><p><strong>Code review became the chokepoint everybody wants to own.</strong> Anthropic took Code Review to general availability with Quick, Standard or Deep set per repository (<a href="https://x.com/ClaudeDevs/status/2110866210573386065">ClaudeDevs</a>). Linear’s founders call Linear reviews “the best code review tool in existence” — their own product, so discount the superlative (<a href="https://x.com/karrisaarinen/status/2107283359890374784">karrisaarinen</a>). Last month Cursor’s Origin was about owning the place everything passes through. Review is that place now: when an agent writes half the PRs, reading is the bottleneck, not writing. Deep on the booking flow, Quick on the long tail. Our squads plan in Linear and review in GitHub, and that default just became a decision.</p><p><strong>Evals got a playbook, and step one is reading, not writing.</strong> Hamel Husain and Shreya Shankar’s evals skill pack for coding agents starts with error discovery on real traces, and makes you validate the LLM judge against human labels before you trust it (<a href="https://x.com/lennysan/status/2102464854594662480">lennysan</a>). Last month claude plugin eval gave skills a regression test. Now the features get one. A suite written before anyone reads the traces tests the failures you imagined, and a judge never calibrated against a human is a second opinion with no track record. Every LLM feature we ship should start with an afternoon of reading transcripts.</p><p><strong>The cheap model in your loop will go quiet one day, so decide now what happens when it does.</strong> Last month I wrote that the cheapest model in your stack decides what the expensive one never sees. TypeSafe’s new Claude Code skill for Jev fixes the mundane failure — agents asking one question per call — and cites a 12.2x bill drop from batching 13 questions into one, a figure from TypeSafe’s own cookbook (<a href="https://x.com/mnilax/status/2103901384135295088">Mnilax</a>). The sharper lesson is in the post he quoted: the checker went silent, and the whole agent stopped (<a href="https://x.com/Mnilax/status/2101015355133227348">Mnilax</a>). Give every checker a default for silence — pass the output tagged unverified, escalate, retry, or wake a human. Every guardrail is also a new way for the pipeline to stop.</p><p><strong>One person now runs enough agents to need management.</strong> Garry Tan gives every bug its own agent thread and relies on Capy’s “virtual standup” to stop them colliding (<a href="https://x.com/garrytan/status/2107527654530576501">garrytan</a>). Danny Postma is building a kanban board on top of Orca (<a href="https://x.com/dannypostma/status/2106215980968136958">dannypostma</a>). Peter Steinberger’s CodexBar now tracks a dozen more AI coding providers from the menu bar (<a href="https://x.com/steipete/status/2102672670689275965">steipete</a>). A standup, a board, a budget: the tools teams invented for coordinating people. Last month I said parallel agents degrade past a handful unless something durable carries state between them. The fix showing up isn’t a smarter model. It’s management.</p><p><strong>And Google just asked to be the only agent you need.</strong> Thomas Kurian announced Gemini as “a new single universal agent for work”: one prompt box, one memory, sub-agents, a “coworker agent” with its own identity, and routing across models (<a href="https://x.com/ThomasOrTK/status/2108245829530325307">ThomasOrTK</a>). We run on Workspace and GCP, so this lands on us directly, and it’s the opposite of the multi-LLM stance we keep defending. Here’s the gut punch. Every launch this month sold a place to put the agent: a cloud box that sleeps, a webhook, a schedule, a channel, a review queue, one prompt box. Not one of them tells you what to say to it. Boris’s 50% isn’t a Slack feature. It’s a person who knew what to hand over, wrote the order once, and stopped watching. Google will sell you a coworker agent. Nobody sells the coworker who knows what to ask it. The gap between your best AI-native person and your median one is still the most important number on your org chart. This month it’s about who can delegate.</p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*dqV8G1ziMnthK4XtXkKvZQ.png" /></figure><h3>AI / Agents</h3><p><strong>Claude Tag in Slack writes over half of Boris Cherny’s PRs</strong> — Proactive, programmable, with memory and, since 24 September, your personal connectors. The pattern to copy is the standing order left in a channel (<a href="https://x.com/bcherny/status/2103538666597691552">bcherny</a>).</p><p><strong>DigitalOcean runs Claude Code and Codex as cloud agents that pause when idle</strong> — Billing stops when the session idles, and it wakes in about 300ms on any device. No pricing or product page in the post, so treat the figures as claims (<a href="https://x.com/VadimStrizheus/status/2102449430771417589">VadimStrizheus</a>).</p><p><strong>T3 Code adds webhook-triggered agent tasks</strong> — An HMAC-checked event goes straight into a task’s prompt template, with up to 24 hours of replay through T3 Connect. Exactly the shape a Linear triage bot needs (<a href="https://x.com/jullerino/status/2107510882435789018">jullerino</a>).</p><p><strong>Google launches Gemini as one universal agent for work</strong> — Sub-agents, a coworker agent with its own role and identity, and routing across models for quality and cost. Google bidding for the enterprise agent layer against Claude and ChatGPT (<a href="https://x.com/ThomasOrTK/status/2108245829530325307">ThomasOrTK</a>).</p><p><strong>Hamel Husain and Shreya Shankar’s evals skill pack</strong> — Error discovery first, then synthetic data, judge prompts, and calibrating the judge against human labels. The authors say it encodes mistakes seen across 50+ companies (<a href="https://x.com/lennysan/status/2102464854594662480">lennysan</a>).</p><h3>Engineering / DevOps</h3><p><strong>Claude Code Review goes GA, with review depth set per repository</strong> — Quick, Standard or Deep per repo on Team and Enterprise. Spend the deep reviews where a regression costs money (<a href="https://x.com/ClaudeDevs/status/2110866210573386065">ClaudeDevs</a>).</p><p><strong>OpenShip 0.8.0 draws your deployment as a live topology</strong> — An interactive service graph, plus Codex and Claude Code operating the platform through MCP. Open-source PaaS in the Coolify category is converging on both (<a href="https://x.com/openshipio/status/2104187404533838043">openshipio</a>).</p><h3>Product Management</h3><p><strong>Linear opens its first European hub in London</strong> — A launch evening on 10 November, from a company that now calls itself “the product development system for teams and agents” (<a href="https://x.com/linear/status/2112193621432717591">linear</a>).</p><h3>Others</h3><p><strong>A seven-day method for winning Google AI Overview citations</strong> — Jake Ward’s process: track one AI Overview for a week, reverse-engineer its citations, build the page around them. The detail sits in an image, so treat it as a method to test. Our SEO team has proposed exactly that (<a href="https://x.com/jakezward/status/2107453596107219294">jakezward</a>).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f67f719c807e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Context Window — September 2026]]></title>
            <link>https://medium.com/@smnbss/context-window-september-2026-f009c4468d2c?source=rss-5de037daf82f------2</link>
            <guid isPermaLink="false">https://medium.com/p/f009c4468d2c</guid>
            <category><![CDATA[ai-sharing]]></category>
            <category><![CDATA[software-factories]]></category>
            <dc:creator><![CDATA[Simone Basso]]></dc:creator>
            <pubDate>Mon, 21 Sep 2026 06:25:56 GMT</pubDate>
            <atom:updated>2026-09-21T06:25:56.796Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>The software factory became a category this month — a name, an architecture, and its first objection.</strong> Zach Lloyd published the case for Warp Factories: laptop-run agents can’t be measured and can’t be governed, so move the whole loop — triage, spec, implement, review, verify — into the cloud under a foreman orchestrator, with the factory itself defined as version-controlled code (<a href="https://x.com/zachlloydtweets/status/2089734641863286877">zachlloydtweets</a>). A day later, the receipts from Warp’s own: 1,000+ PRs in one week across 8+ repos, over a third merged, roughly a 6-hour cycle time (<a href="https://x.com/zachlloydtweets/status/2090131176078098700">zachlloydtweets</a>). Josh Rosen supplied the definition — “if an agent harness is an opinionated agent framework, a software factory is an opinionated orchestrator” — plus six patterns shared by Factory.ai, Warp and Vercel’s Foreman (<a href="https://x.com/josharosen/status/2094076039194358070">josharosen</a>). Google shipped the same thesis as prompts inside whichever coding agent you already use (<a href="https://x.com/akshay_pachaar/status/2089707367172882669">akshay_pachaar</a>). The objection arrived from David Crawshaw via Dex Horthy: these need to be open systems you compose and own, not turnkey stacks (<a href="https://x.com/dexhorthy/status/2093757824849301669">dexhorthy</a>). He’s right, and it’s premature. The governance half is what lands on us today — agents on individual machines, no shared skill set, no data exhaust. That isn’t a tooling gap. It’s an unmeasured production system.</p><p><strong>The agent moved out of the terminal, and the fight is over which room it lands in.</strong> Slack demoed Slack Code, where a GitHub-connected agent gets its own channel per repository and the team watches, comments and steers it in real time (<a href="https://x.com/slackhq/status/2090870662382424338">slackhq</a>). Peter Steinberger put the principle in one line: a CLI is fine for solo work, but a UI with visualisations, surfaced where your team already works, is the better default (<a href="https://x.com/steipete/status/2091650136506327253">steipete</a>). Claude, meanwhile, started sending Gmail replies and managing Drive files directly, with the user deciding when approval is required (<a href="https://x.com/claudeai/status/2089806039088517356">claudeai</a>). Last month I said the durable idea in Jack Dorsey’s BUZZ was a shared channel where every message is reviewable. Slack just shipped it with real distribution behind it. The terminal was never the point — it was only the place the work was legible. Once a team can see it, review stops being optional. Which is exactly why “you control when it needs your approval” deserves a hard look before anyone switches on an agent that sends mail.</p><p><strong>Skills became a company asset — and then, finally, got a test.</strong> Notion made itself the authoring home, syncing a team’s skills into Claude Code, Codex, Cursor, Gemini or Grok as a SKILL.md plus approved files (<a href="https://x.com/notionhq/status/2090829567451861122">notionhq</a>). Anthropic took Computer Use, the browser tool, the Skills API and the Files API to GA, positioning versioned skills as the substrate under Managed Agents rather than a convenience (<a href="https://x.com/claudedevs/status/2090540270219567575">claudedevs</a>). Google&#39;s WikiSkill paper argued for the layer beneath all of it: consolidate execution traces into a persistent wiki first, then evolve skills against that wiki. The ablations carry the argument — the wiki holds most of the gain, and skills evolved by one model transfer across model families (<a href="https://x.com/dair_ai/status/2093324233158045788">dair_ai</a>). Then claude plugin eval shipped: score your skill against test cases, then re-run each case <em>without</em> it to measure what it actually contributed (<a href="https://x.com/trq212/status/2098531560643539440">trq212</a>). Last month I wrote that agent config is the new tech debt. This is the first tool that lets you price it. Our own corpus has no regression harness at all, so a model upgrade changes behaviour silently and we find out when a rebuild writes a wrong page. That delta is also the only honest answer to &quot;is this skill earning its context budget?&quot; Several of ours are not.</p><p><strong>The integration layer stopped being a per-user chore.</strong> Cloudflare previewed MCP v2, a rewritten stateless core built to run natively on Workers (<a href="https://x.com/fayazara/status/2091395496262930886">fayazara</a>). Anthropic took enterprise-managed authentication for MCP connectors to GA, so an admin authorises centrally through the identity provider and nobody runs an OAuth dance per connector (<a href="https://x.com/ClaudeDevs/status/2091953609185657251">ClaudeDevs</a>). Marc Köhlbrugge asked what everyone past a handful of servers is now asking — is anyone actually running one gateway in front of all their MCPs (<a href="https://x.com/marckohlbrugge/status/2092235834401419361">marckohlbrugge</a>) — and Uber Engineering answered by publishing the harness and gateway it built for itself (<a href="https://x.com/UberEng/status/2093444169037762840">UberEng</a>). WebMCP landed in ChatGPT, so a site can expose its own callable tools instead of being clicked through (<a href="https://x.com/wesbos/status/2092349927129035086">wesbos</a>). Last month “MCP went stateless” was the headline. This month it’s plumbing: authorised centrally, aggregated behind a gateway. We register a growing MCP surface client by client, which is precisely the shape Uber moved off.</p><p><strong>The company brain got a price tag, and it’s far below what anyone budgeted.</strong> Claire Vo’s episode with Sharadh from Stripe walks through Kai, their internal brain: 1.5 engineers, two weeks, working. Three pieces are worth stealing outright — projects as the governance unit, skill routing with telemetry so you can see which skills actually get invoked and at what cost, and a skills platform built to scale to 10,000 teammates (<a href="https://x.com/clairevo/status/2097105221948436500">clairevo</a>). Garry Tan made the portability half explicit: gbrain runs against any harness and any hosted Postgres with pgvector, not only a local install (<a href="https://x.com/garrytan/status/2089733961492930659">garrytan</a>). And the loudest post of the month on the subject was a workshop funnel — 50 bots on a $200 plan, $1.2M on a headcount of one, no artifact attached, every figure asserted by the man selling the workshop (<a href="https://x.com/kingwilliam_/status/2098574704797741267">kingwilliam_</a>). Discount the numbers entirely and the thesis still holds, because it’s the bet this whole stack is: parallel agents degrade past a handful unless something durable carries state between them. Last month I called the Cerebras write-up a reference implementation of our own gbrain bet. Stripe supplied the cost, and named the two pieces we don’t have — routing telemetry, and a governance boundary that isn’t “whatever the agent happens to be able to reach”.</p><p><strong>The cheapest model in your stack is the one that decides what the expensive one never sees.</strong> Spotify cut Claude Code token usage by 90%, and the mechanism is not a better prompt. Most of what a coding assistant does carries no judgment — it opens five files to answer a question about one — so two cheap sub-assistants take that work and the expensive model never sees the payload. The decisive detail is enforcement: written instructions were ignored, so anything over 350 lines is now stopped before it opens and routed to the cheap model. “Written rules are a suggestion, a block is not” (<a href="https://x.com/undefinedKi/status/2095942506433089832">undefinedKi</a>). Tobi Lütke reported a 0.8-billion-parameter fine-tune beating a frontier model on one specialised task, with the precondition being process rather than size (<a href="https://x.com/tobi/status/2094808564355191249">tobi</a>). Then the Jev results landed together: a session compacted from nearly 1M tokens to 86K in about a second (<a href="https://x.com/altryne/status/2100739055923425589">altryne</a>), 712 of 948 eventual churners flagged from three journeys each for six cents (<a href="https://x.com/regalstreak/status/2101189571375493239">regalstreak</a>), and an SEO and GEO run falling from roughly $250 to roughly $25 on model substitution alone (<a href="https://x.com/HamonPaulm/status/2101265909826609278">HamonPaulm</a>). Label those last two honestly — vendor-reported, unaudited, attached to early-access pitches. The direction isn’t debatable anyway. Every one of these takes the mechanical 80% off the frontier model, and the version that survives contact with a real team is the one enforced in a gate rather than written in a rulebook. We proxy every LLM call for cost monitoring and we route almost none of it. That’s the gap, and it’s ours.</p><p><strong>And the metering layer got bought.</strong> Last month I wrote that Stripe was reportedly in talks to acquire OpenRouter for around $10B, and closed on “the app layer is being commoditised, the metering layer is being acquired”. Bloomberg reported on 16 August that the agreement is finalised, at more than $7B (<a href="https://x.com/business/status/2089077253778915636">business</a>). A payments company paid over seven billion dollars for the thing that sits in the request path and sees every token, every provider price and every fallback. In the same fortnight Cursor launched Origin, its own code-hosting platform, opening the onboarding with “sync your repos from GitHub” (<a href="https://x.com/cursor_ai/status/2089399057659596847">cursor_ai</a>). Same move, different layer: own the place everything has to pass through. Here’s the gut punch. Every theme above this one is an infrastructure story — factories, control surfaces, skill libraries, gateways, company brains, routing — and there is not one of them you cannot buy. The hardest thing I read all month had no AI in it at all: Luca Ferrari on the Bending Spoons operating model, where hiring became a science, there are no bonuses, there are no job titles, and everyone in the company holds the same job (<a href="https://x.com/davidsenra/status/2099136019572666507">davidsenra</a>). A company that redesigned how people are organised, then scaled by acquisition. The gap between your best AI-native person and your median one is still the most important number on your org chart. What changed this month is that I can no longer pretend a purchase closes it. Every vendor above will sell you the factory. Nobody will sell you the organisation that can run one.</p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*dqV8G1ziMnthK4XtXkKvZQ.png" /></figure><h3>AI / Agents</h3><p><strong>Warp Factories: cloud software factories pitched as infrastructure you own</strong> — The long-form argument starts from two failures of laptop-run agents: you can’t measure coding-agent ROI as usage climbs, and you can’t govern it when every engineer installs a bespoke agent with access to everything they’re logged into. The answer is triage, spec, implement, review and verify agents in the cloud under a foreman orchestrator, with factories defined as version-controlled code, and explicit multi-harness support so Claude Code or Codex can be the worker (<a href="https://x.com/zachlloydtweets/status/2089734641863286877">zachlloydtweets</a>).</p><p><strong>Warp’s own factory: 1,000+ PRs in a week, ~6-hour cycle time</strong> — The usage data behind the pitch, posted a day later: 8+ repos, 1,000+ opened PRs in the last week with over a third merged, 5,000+ total runs. The most concrete external benchmark yet for cloud-agent throughput at a real engineering org, and the number to hold any “we’re doing agentic engineering” claim against (<a href="https://x.com/zachlloydtweets/status/2090131176078098700">zachlloydtweets</a>).</p><p><strong>“A harness is an opinionated framework, a factory is an opinionated orchestrator”</strong> — Josh Rosen’s one-line distinction, backed by an architecture essay on Factory.ai, Warp and Vercel’s Foreman. Six shared patterns: the factory owns the outer loop, workers are ephemeral and isolated per stage, state lives outside the agent, verification is a first-class stage, and full autonomy is still aspirational. His closing argument is the sharp one — a strong orchestration layer commoditises the coding agent underneath it (<a href="https://x.com/josharosen/status/2094076039194358070">josharosen</a>).</p><p><strong>Google’s Agents CLI collapses the agent lifecycle into the coding agent you already use</strong> — Setup, build, deploy, govern, evaluate and publish become plain-English prompts rather than six separate consoles. Govern is the stage most self-built setups skip entirely: it provisions a least-privilege identity, screens for prompt injection, and confines the agent to an egress allow-list (<a href="https://x.com/akshay_pachaar/status/2089707367172882669">akshay_pachaar</a>).</p><p><strong>Multica: an open-source board where 20 coding-agent CLIs get assigned issues like teammates</strong> — Self-hostable, pitched by its author as “a harness for managing all your harnesses”. You assign an issue, the agent runs on a runtime you control so code never leaves your machine, comments as it goes, raises blockers and hands back a diff. Plus Squads that put agents and humans on one team, and reusable Skills playbooks (<a href="https://x.com/jiayuan_jy/status/2089297182700847505">jiayuan_jy</a>).</p><p>claude plugin eval<strong> — regression tests for skills across model releases</strong> — Create test cases, run the skill against them, score the runs, then re-run each case without the skill to measure the delta. The framing is drift: people can&#39;t tell whether their skills still work after a new model lands. That last step is what makes it more than a smoke test — it measures marginal contribution, which is also the only honest way to decide whether a skill earns its context budget (<a href="https://x.com/trq212/status/2098531560643539440">trq212</a>).</p><p><strong>Notion syncs skills straight into Claude Code, Codex, Cursor, Gemini and Grok</strong> — Author a skill in Notion as the team’s source of truth, then download it into a local coding agent as a SKILL.md plus approved supporting files. The second major vendor after Anthropic&#39;s own Skills API to treat a skill as a portable, tool-agnostic artifact. The bet underneath is that the authoring surface for skills won&#39;t be git (<a href="https://x.com/notionhq/status/2090829567451861122">notionhq</a>).</p><p><strong>WikiSkill: consolidate experience into a persistent wiki before you evolve the skills</strong> — A Google paper separating raw execution traces, an accumulated-knowledge wiki, and the executable skills themselves. The ablations carry the argument: the wiki accounts for most of the gain, smaller models with evolved skills beat substantially larger models without them, and skills evolved by one model transfer across model families — sometimes beating skills a model evolved for itself (<a href="https://x.com/dair_ai/status/2093324233158045788">dair_ai</a>).</p><p><strong>Computer Use, the browser tool, the Skills API and the Files API all go GA</strong> — Four platform capabilities out of preview together, letting Claude automate work inside applications that have no API, with versioned skills plus reusable files positioned as the way you build Managed Agents. This is the line between a computer-use demo and something you can build a reliable automation on — relevant to any workflow blocked on a vendor UI with no API behind it (<a href="https://x.com/claudedevs/status/2090540270219567575">claudedevs</a>).</p><p><strong>Stripe built “Kai”, a company brain, with 1.5 engineers in 2 weeks</strong> — Claire Vo’s episode with Sharadh covers Stripe’s internal AI platform. Three architectural pieces are named: projects as the governance unit, skill routing plus telemetry so you can see which skills get invoked and at what cost, and a skills platform designed for 10,000 teammates. The cost figure is the headline; the routing telemetry is the part most self-built brains are missing (<a href="https://x.com/clairevo/status/2097105221948436500">clairevo</a>).</p><p><strong>GBrain is harness-agnostic, and works on any hosted Postgres with pgvector</strong> — Garry Tan clarifies the memory layer runs against Grok Bot, Claude Code, Codex, Hermes Agent, OpenClaw and OpenCode, and against any hosted Postgres offering pgvector rather than only a local install. The framing is ownership — not coupling your agent’s memory and skills to one vendor. For anyone already running this configuration, the “any hosted Postgres” half is the door to getting the index off a single laptop (<a href="https://x.com/garrytan/status/2089733961492930659">garrytan</a>).</p><p><strong>Spotify cut Claude Code token usage by 90% by blocking the expensive model from cheap work</strong> — Two cheap sub-assistants absorb the no-judgment volume: one opens files and returns a summary, the other writes repetitive code from an example, and the expensive model never sees either payload. Enforcement is the decisive detail — written instructions were ignored, so anything over 350 lines is now stopped before it opens. Two things stayed expensive: edits still need the real file, and the cheap worker missed a bug the expensive model caught in seconds (<a href="https://x.com/undefinedKi/status/2095942506433089832">undefinedKi</a>).</p><p><strong>Shopify’s 0.8B fine-tune beats a frontier model on one narrow task</strong> — Tobi Lütke names the precondition as process rather than size or data volume: tiny special-purpose models work “incredibly well” only with a strong self-improving recursive flywheel around them. The benchmark is an attached image, so exact scores aren’t readable from the post. Still the strongest recent public data point for small-specialist-beats-general on a bounded task — and most high-volume pipeline work is exactly that shape (<a href="https://x.com/tobi/status/2094808564355191249">tobi</a>).</p><p><strong>A Jev-based plugin compacts a 1M-token session to 86K in about a second</strong> — fast-jev-compaction uses a small model inside Claude to review a session&#39;s tool calls and strip the unnecessary ones. A roughly 92% context reduction at sub-second cost changes the economics of long agent sessions, not just their comfort. The evidence is two screenshots and one person&#39;s run, so treat the ratio as indicative — but the mechanism, a cheap model pruning an expensive model&#39;s history, is the same pattern as the Spotify result (<a href="https://x.com/altryne/status/2100739055923425589">altryne</a>).</p><p><strong>Claude now sends Gmail replies and manages Drive files directly</strong> — Ask it to reply to a thread and it drafts and sends, with the user controlling when approval is required, on all paid plans. This pushes agentic write actions past drafting into acting on your behalf, which is precisely the class of action permission gates exist to bound. The default scoping of “you control when it needs approval” is the thing to inspect before rollout, not after (<a href="https://x.com/claudeai/status/2089806039088517356">claudeai</a>).</p><p><strong>Slack Code turns a channel into a live, multiplayer coding session</strong> — A GitHub-connected agent spins up a channel tied to one repository, then works on a fix while the team watches, comments and steers in real time. A normally solo terminal session becomes shared and observable inside chat infrastructure everyone is already in — another entrant arguing that chat, not the IDE, is where agentic work gets reviewed (<a href="https://x.com/slackhq/status/2090870662382424338">slackhq</a>).</p><h3>Engineering / DevOps</h3><p><strong>Cloudflare previews MCP v2: a stateless, Workers-native core</strong> — A rewritten core built to run natively on Workers, plus protocol upgrades, a new feature lifecycle and an SDK migration path. For anyone operating remote MCP servers, a stateless core changes the deployment story outright — no sticky sessions, horizontal scaling like any other service (<a href="https://x.com/fayazara/status/2091395496262930886">fayazara</a>).</p><p><strong>Enterprise-managed auth for MCP connectors goes GA</strong> — Admins get centralised authorisation through their identity provider and users get automatic tool and data connections with no per-connector OAuth flow. That setup cost was the main onboarding tax on any org-wide MCP rollout, so it’s worth revisiting any MCP surface you shelved because the per-person effort was too high (<a href="https://x.com/ClaudeDevs/status/2091953609185657251">ClaudeDevs</a>).</p><p><strong>Uber published the harness and MCP gateway it built for itself</strong> — An X-native article, so the body isn’t retrievable through the API, but the subject is confirmed and a companion video covers it. The relevance is timing: it’s the direct answer to the gateway question circulating the same week, from a company at a scale where registering servers client by client stopped working long ago (<a href="https://x.com/UberEng/status/2093444169037762840">UberEng</a>).</p><p><strong>WebMCP lands in ChatGPT</strong> — Websites can expose their own callable tools to an agent instead of being scraped and clicked through. Wes Bos notes the obvious near-term move is exposing Search itself as a WebMCP tool. This takes WebMCP from spec to real client support, which makes “what tools does our site publish to agents?” a live product question (<a href="https://x.com/wesbos/status/2092349927129035086">wesbos</a>).</p><p><strong>Cursor ships Origin, its own code-hosting platform</strong> — Live now, with onboarding that begins by syncing your repos from GitHub. 4.6M views, the loudest post in the window. It moves Cursor from an editor sitting on someone else’s git host to owning the repository — where reviews, CI hooks and the agent’s view of history all live. Market signal rather than something most teams will act on (<a href="https://x.com/cursor_ai/status/2089399057659596847">cursor_ai</a>).</p><h3>Design</h3><p><strong>Claude Code ships a </strong>/design<strong> skill — editable artboards in the CLI, then implementation</strong> — In research preview. Run /design, get editable artboards for your UI, pick one, tweak it, then have Claude implement it as code. The coding agent is absorbing the design-handoff step, collapsing &quot;Figma → ticket → implementation&quot; into one session — and quietly invalidating any review process that assumes a separate design artifact exists to review (<a href="https://x.com/ClaudeDevs/status/2089471692762673408">ClaudeDevs</a>).</p><h3>Product Management</h3><p><strong>A UI beats a CLI for team-visible agent work</strong> — Peter Steinberger’s argument is short and pointed, and it comes from someone actively building agentic tooling. The product consequence is the one to sit with: if an agent’s work is only legible in a terminal, exactly one person can review it (<a href="https://x.com/steipete/status/2091650136506327253">steipete</a>).</p><p><strong>Jev flags 712 of 948 churning users from three journeys each, for $0.06</strong> — Shown only each user’s first three journeys with the outcome hidden, the model returned in about 10 seconds a prediction of who would quit, a confidence figure, and where each journey went wrong. A second pass drafted a fix for the recurring issue. It’s an early-access pitch, so accuracy and cost are vendor-reported. The shape is what matters — churn prediction plus root-cause drafting at a price where running it on every cohort is effectively free (<a href="https://x.com/regalstreak/status/2101189571375493239">regalstreak</a>).</p><h3>Others</h3><p><strong>Stripe finalises the OpenRouter acquisition at more than $7B</strong> — Bloomberg reported on 16 August that the agreement is done. OpenRouter’s whole product is one API that lets a company switch between models from different providers without rewriting the integration. A payments company paying over seven billion for a model router says the metering and billing layer of AI traffic is the valuable part, not the models. For anyone routing through it, pricing, terms and data-handling posture are now a new owner’s decision (<a href="https://x.com/business/status/2089077253778915636">business</a>).</p><p><strong>Luca Ferrari on the Bending Spoons operating model: no titles, no bonuses, hiring as a science</strong> — A long-form conversation mapping an explicit operating model rather than a founder story: hiring turned into a science, no bonuses, everyone holding the same job, no job titles, and a proprietary internal operating system underneath it all. One of the very few detailed public accounts of a European company that removed titles, bonuses and role boundaries and still scaled through acquisition. Note the post includes paid partnerships (<a href="https://x.com/davidsenra/status/2099136019572666507">davidsenra</a>).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f009c4468d2c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Context Window — August 2026]]></title>
            <link>https://medium.com/@smnbss/context-window-august-2026-924ec9755e09?source=rss-5de037daf82f------2</link>
            <guid isPermaLink="false">https://medium.com/p/924ec9755e09</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai-sharing]]></category>
            <dc:creator><![CDATA[Simone Basso]]></dc:creator>
            <pubDate>Mon, 17 Aug 2026 08:47:53 GMT</pubDate>
            <atom:updated>2026-08-17T08:47:53.801Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Agents stopped answering and started signing in.</strong> xAI put Grok Bot into early beta and didn’t pitch it as an assistant — it signs in to your existing tools, operates them the way a person would, and hands back finished work (<a href="https://x.com/bot/status/2087224798078517251">bot</a>). Andrew Ng and Rohit Prasad launched OpenWorker on the same premise, open source and local, delivering a polished document or a sent Slack message rather than a suggestion, and checking in before anything consequential (<a href="https://x.com/andrewyng/status/2080333504446108104">AndrewYNg</a>). Linear shipped Loops so its agent runs recurring workflows on a schedule instead of waiting to be asked (<a href="https://x.com/linear/status/2079233260161323371">linear</a>). Claude Managed Agents got per-agent effort levels, 500 skills per session, webhooks and sub-agent event streaming (<a href="https://x.com/claudedevs/status/2080009523952263295">ClaudeDevs</a>). The unit of delivery moved from a response to a completed task. And “signs in to your tools like you do” is not a feature description — it’s a credentials posture. Last month I wrote that agent identity is the boring plumbing that makes this real. This month everyone shipped the capability and left the plumbing to you.</p><p><strong>The security layer for agents arrived before most companies thought to ask for it.</strong> Perplexity open-sourced Numbat, a harness-agnostic agent detection and response layer for macOS, Linux and Windows that blocks selected actions <em>before</em> they execute rather than reviewing logs afterwards (<a href="https://x.com/perplexity_ai/status/2082511900580196596">perplexity_ai</a>). OpenAI quietly shipped an open-source Codex Security CLI — Hacker News found it before the announcement — that tracks findings across runs and verifies that fixes actually closed something (<a href="https://x.com/OpenAI/status/2082263717916586117">OpenAI</a>). Anthropic put a Claude Security plugin into beta so you scan changes pre-commit from the terminal (<a href="https://x.com/claudeai/status/2079990597973057691">claudeai</a>). And Semantica open-sourced the Palantir-shaped part nobody enjoys building: context graphs, explainable reasoning, permanent audit trail (<a href="https://x.com/tom_doerr/status/2083434480723718478">tom_doerr</a>). Four credible entrants in three weeks, in a category that barely existed in spring. Vendors have already priced in the fact that agents will run on employee laptops holding real credentials. Most orgs haven’t. The distance between “we rolled out agents” and “we can see what they did” is exactly where the first bad incident lives.</p><p><strong>Somebody finally tried to build the room where humans and agents actually work together.</strong> Jack Dorsey launched BUZZ — open-source groupchat for teams of people <em>and</em> agents, model-agnostic and decentralized, explicitly aimed at reducing dependency on Slack and GitHub (<a href="https://x.com/jack/status/2079605800998146171">jack</a>). ThomPete’s launch note filled in the substance: built at Block, every team gets its own git server that humans and agents fully control, and the UI was sculpted over time by agents rather than designed up front (<a href="https://x.com/hello_world/status/2079645150276464643">Hello_World</a>). Then came the demo that made the case better than either announcement — three Hermes agents on three different machines, a local PC, a Spark box over SSH and a VPS, all joined to one Buzz group as separate members, one researching and one cross-checking its claims and issuing a correction, with no human copy-pasting between terminals (<a href="https://x.com/IBuzovskyi/status/2084587252261294108">IBuzovskyi</a>). Anthropic pushed the same idea sideways: Claude in Chrome sessions now carry to desktop, web and mobile, with skills and connectors working inside the browser (<a href="https://x.com/claudeai/status/2087635262390026525">claudeai</a>). Ignore the Slack-killer framing. The two durable ideas here are an agent that remembers you across surfaces, and a shared channel where every message is signed and reviewable. Also worth holding onto: that three-machine demo’s headline failure was a missing binary on PATH and an error message that lied about the cause. That’s where this genuinely is.</p><p><strong>The company brain went from thesis to published blueprint.</strong> Cerebras documented the full build of its internal knowledge base — now fielding more than 15,000 questions a day from humans, automations <em>and</em> agents, three months after launch (<a href="https://x.com/cerebras/status/2077822555159945507">cerebras</a>). The design choices are the valuable part. They refused to make anyone migrate their data, extracting from Slack, GitHub, docs and Jira into a single Postgres embeddings table. Retrieval is four scorers fused at query time with Reciprocal Rank Fusion: full-text for literal strings like error messages and hostnames, embeddings for paraphrase, IDF to kill filler (“sounds good, thanks!” sits near every query in embedding space), and explicit age decay because a six-month-old Slack answer often describes infrastructure that no longer exists. Best idea in the whole post: they distil each thread into a question, summary and resolution with an LLM and embed <em>that</em>, not the raw transcript — and it measurably raised accuracy. Replit arrived at the same place from the other end, with a self-correcting semantic layer spanning databases, docs and conversations so anyone can ask questions that used to take data scientists weeks (<a href="https://x.com/amasad/status/2084415670486499779">amasad</a>). I said last month this was the bet behind our own gbrain stack. Now there’s a reference implementation that independently landed on Postgres, hybrid retrieval and MCP — and it hands you four cheap upgrades. We’re stealing all four.</p><p><strong>The plumbing got boring, which is precisely the milestone.</strong> MCP shipped its 2026-07-28 revision, billed as the biggest change since the protocol launched: it&#39;s stateless now, so remote servers stop needing sticky sessions and start scaling horizontally like any other service (<a href="https://x.com/ClaudeDevs/status/2082164248697069935">ClaudeDevs</a>). LangChain put Managed Deep Agents into beta, where you write an agent as a <em>directory of files</em> — instructions, skills, tools, MCP connectors, middleware, memory, schedules and evals each get a slot — and LangSmith runs the harness and keeps sessions alive across restarts (<a href="https://x.com/caspar_br/status/2088345102540587356">caspar_br</a>). Anthropic went the opposite way and cut roughly 80% of the Claude Code system prompt for its newest models, because capable models need less hand-holding, not more scaffolding (<a href="https://x.com/trq212/status/2080710971228918066">trq212</a>). Three signals, one direction: the agent is a folder of markdown plus tools, the runtime is somebody else&#39;s problem, and the prompt is shrinking. Now the bill. Trevin Chow named the thing anyone running multiple harnesses already feels — rules, skills, hooks and accumulated corrections sprawl across Claude Code, Codex and Cursor and form a dependency graph nobody can see (<a href="https://x.com/trevin/status/2084679477842608223">trevin</a>). We&#39;ve been bitten by exactly that, twice. Agent config is the new tech debt.</p><p><strong>The build-versus-buy line moved again, and it moved fast.</strong> Comp AI got quoted $80k a year by HubSpot for a sales team of fewer than ten people, spent half a day vibe-coding a replacement, and open-sourced it under MIT — agentic-first meant literally, with durable research agents as the core primitive rather than a pipeline app with an AI sidebar bolted on (<a href="https://x.com/lewiscarhart/status/2083610805069611230">lewiscarhart</a>). In the same stretch one developer shipped quill, a local private replacement for Granola, and parrot, on-device Whisper dictation instead of a Wispr Flow subscription — both tiny macOS menu-bar apps with no accounts and no upload step (<a href="https://x.com/dremnik/status/2081879207509217473">dremnik</a>). And Hetzner started handing out a free, explicitly no-SLA inference API for DeepSeek v4, GLM 5.2 and Kimi (<a href="https://x.com/maxjendrall/status/2087113184423333939">maxjendrall</a>). For the single-user case, a weekend now credibly displaces a funded SaaS product. The moat moves to sync, team features and accuracy at the tail — the parts a weekend doesn’t buy. But hold the counterweight from the same month: Stripe is reportedly in talks to buy OpenRouter for around $10B (<a href="https://x.com/polymarketmoney/status/2080466494832497004">PolymarketMoney</a>). The app layer is being commoditised. The metering layer is being acquired.</p><p><strong>And the number that should bother you is 625.</strong> Olivia Moore, on Ramp card-spend data: the median company spends $12 per employee per month on AI. The top 1% spend $7,500 (<a href="https://x.com/omooretweets/status/2088314089659662799">omooretweets</a>). Caveat it honestly — that’s Ramp’s customer base, not the market, and a per-employee average says nothing about how concentrated the spend is inside those top firms. We sit far above the median and nowhere near the ceiling. Here’s the gut punch. Last month I said cost had stopped being a finance problem and become an architecture decision. This data says something harsher, because nobody in that thread demonstrates that spending more produces more. A 625× spread isn’t a pricing story. Twelve dollars a month is one seat and a good intention. Seventy-five hundred is a company that rebuilt how it works. The gap between your best AI-native person and your median one is still the most important number on your org chart — and it did not close this month.</p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*dqV8G1ziMnthK4XtXkKvZQ.png" /></figure><h3>AI / Agents</h3><p><strong>Grok Bot enters early beta as an “AI teammate” that signs into your tools</strong> — xAI’s pitch is that it operates your existing software the way a person would and returns completed work, not suggestions, putting it in the computer-use/autonomous-work category rather than the chatbot one it started in. Enormous distribution for a product launch (~15.3M views), and details at x.ai/bot are still gated, so integrations, autonomy limits and pricing aren’t publicly verifiable. The part that matters for anyone evaluating it: “signs in to your tools” is a question about which accounts an agent holds and what it can approve (<a href="https://x.com/bot/status/2087224798078517251">bot</a>).</p><p><strong>OpenWorker: Andrew Ng’s open-source AI coworker that hands over finished deliverables</strong> — Built with Rohit Prasad, it works across your files and everyday tools to produce a polished document, send a Slack message or untangle a calendar, checking in before anything consequential. Runs locally on macOS (Windows coming), model-independent via your own API key — GPT 5.6 Sol, Claude Fable, Gemini 3.6, or Kimi/GLM/DeepSeek through Ollama for fully local data — and your data stays on the machine except through providers you pick. The clearest open benchmark yet for what deliverable-first agent UX should look like (<a href="https://x.com/andrewyng/status/2080333504446108104">AndrewYNg</a>).</p><p><strong>Linear introduces Loops — recurring workflows its agent runs on a schedule</strong> — Rather than firing a one-off agent task, a Loop schedules the built-in Linear Agent to execute a defined workflow repeatedly: triage sweeps, status updates, recurring reporting. It moves the agent from reactive assistant to standing, cron-like operator owned by the tool itself. For anyone already running Linear plus coding agents, this is the issue tracker quietly becoming an automation runtime (<a href="https://x.com/linear/status/2079233260161323371">linear</a>).</p><p><strong>MCP </strong><strong>2026-07-28 goes stateless — the biggest protocol change since launch</strong> — Servers no longer need to hold per-session state, which makes remote MCP servers substantially easier to deploy, load-balance and scale horizontally. ~1.5M views within hours, which tells you how much of the ecosystem now sits on MCP as its integration layer. For anyone running always-on remote MCP servers, this removes the sticky-session and reconnect complexity you&#39;re currently working around (<a href="https://x.com/ClaudeDevs/status/2082164248697069935">ClaudeDevs</a>).</p><p><strong>Cerebras publishes the full build of a knowledge base answering 15,000 questions a day</strong> — Three months after launch, serving humans, automations and agents. No migration: they extract from Slack, GitHub, docs and Jira into one Postgres embeddings table behind a uniform query interface. Retrieval fuses four scorers with Reciprocal Rank Fusion — full-text for literal tokens, embeddings for paraphrase, IDF to suppress filler, age decay for stale answers. Slack threads are LLM-distilled into question/summary/resolution and <em>that</em> gets embedded, not the transcript, which raised accuracy measurably. Repos over 40 GB are chunked class → method → block via CocoIndex. The most complete public prior art on company brains (<a href="https://x.com/cerebras/status/2077822555159945507">cerebras</a>).</p><p><strong>Replit built a self-correcting shared semantic layer over databases, docs and conversations</strong> — Everything queryable and joinable regardless of source, so anyone can ask questions that “previously took a team of data scientists weeks.” The framing is the point: when an agent starts with shared context, more people get reliable answers. The differentiator versus a normal serving layer is that it maintains and corrects itself rather than relying on humans to keep models and definitions in sync — which is the actual failure mode of every self-serve analytics ambition (<a href="https://x.com/amasad/status/2084415670486499779">amasad</a>).</p><p><strong>LangChain ships Managed Deep Agents — write the agent as a folder, let LangSmith run the harness</strong> — In beta. You build an agent as a directory of files where instructions, skills, tools, MCP connectors, middleware, memory, sandboxes, identity, channels, schedules and evals each get a documented slot; LangSmith supplies the Deep Agents harness (plan, call tools, manage a filesystem, delegate to subagents) plus a managed runtime that keeps sessions alive across restarts. No servers to run. It’s the same shape as a hand-rolled skills tree, productised — a useful reference point for any build-vs-buy call on agent infrastructure (<a href="https://x.com/caspar_br/status/2088345102540587356">caspar_br</a>).</p><p><strong>Claude Managed Agents adds per-agent effort levels, 500 skills, and webhooks</strong> — You can now tune effort per agent, seed sessions with events, attach up to 500 skills per session, wire webhooks for environments and memory stores, and stream events from sub-agents. That’s granular cost/latency control plus real integration surface. The honest caveat from our own read: the business-side UX still isn’t business-friendly, so “agent for the business” remains an engineering-mediated story for now (<a href="https://x.com/claudedevs/status/2080009523952263295">ClaudeDevs</a>).</p><p><strong>Anthropic cut ~80% of the Claude Code system prompt for its newest models</strong> — Thariq’s takeaway: more capable models need far less hand-holding, so scaffolding that was necessary a generation ago is now counterproductive bloat. A first-party steer on authoring agent instructions — lean on model defaults, spend prompt real estate only on genuinely non-obvious context, treat skills and CLAUDE.md as targeted overrides rather than exhaustive rulebooks. If your agent config has been growing every month, that’s the signal to reverse it (<a href="https://x.com/trq212/status/2080710971228918066">trq212</a>).</p><p><strong>jack launches BUZZ, an open-source groupchat for people and agents</strong> — Pitched as model-agnostic, decentralized, self-sovereign and open source, explicitly built to reduce dependency on Slack and GitHub. It targets the emerging need for shared spaces where humans and agents coordinate as first-class participants, and fits Dorsey’s long-running protocol-over-platform stance. Worth watching as a signal that agent-native collaboration may fragment away from incumbent SaaS rather than being absorbed by it (<a href="https://x.com/jack/status/2079605800998146171">jack</a>).</p><p><strong>Buzz, from the builder’s side: a git server per team, and a UI sculpted by agents</strong> — ThomPete’s launch detail on the same platform, built at Block to replace the loneliness of working solo with agents. Each team gets its own git server that both humans and agents fully control, and the entire platform — including the interface, shaped over time by agents rather than designed up front — was built with agents. Treating agents as teammates with their own infrastructure, not assistants with a chat box, is the substantive claim here (<a href="https://x.com/hello_world/status/2079645150276464643">Hello_World</a>).</p><p><strong>Three agents, three machines, one channel — and the gotcha that breaks it</strong> — A live demo with Hermes agents on a local PC, a remote Spark box over SSH and a VPS, all joined to one private Buzz group as separate members: one researches, the second reads the thread, cross-checks and issues a correction. No human relaying between terminals. Runs through the native gateway path, which unlocks cron delivery, reactions, approvals, memory and session management. The failure mode is the useful part: without the Buzz CLI on PATH the gateway connects, the agent visibly thinks, the reply never lands — and the error blames channel membership (<a href="https://x.com/IBuzovskyi/status/2084587252261294108">IBuzovskyi</a>).</p><p><strong>Claude in Chrome sessions now sync to desktop, web and mobile — with skills and connectors in-browser</strong> — Conversations started in the extension continue on other surfaces, and the more consequential half is that skills and connectors now run inside the browser rather than only in the desktop and web clients. Staged rollout: Max and Team first, Pro “in the coming weeks.” The security consequence is immediate — a skill that reaches internal systems is now reachable from a tab, so agent access control becomes a browser-context question, not just a desktop one (<a href="https://x.com/claudeai/status/2087635262390026525">claudeai</a>).</p><p><strong>Claude Cowork lets you teach a skill by screen recording</strong> — Record yourself doing a task, narrate as you go, and Claude turns the recording into a reusable skill it can run again; it lives under “Record a skill” in the desktop app’s + menu, on Pro, Max and Team. No skill-file authoring, just demonstrate the workflow once. This is demonstration-based agent programming, and it’s the most plausible path yet to non-technical people codifying their own repeatable procedures (<a href="https://x.com/claudeai/status/2079595988998554047">claudeai</a>).</p><p><strong>Claude Code opens an iOS simulator in a panel next to the conversation</strong> — Build and run your iOS app on the desktop app without context-switching to Xcode, so the agent iterates on UI and you see the result immediately. Shipped in public beta, ~1.1M views. It tightens the edit-build-run loop for mobile and pushes Claude Code further from terminal tool toward full IDE-adjacent environment (<a href="https://x.com/claudedevs/status/2079674432038248611">ClaudeDevs</a>).</p><p><strong>Comp AI open-sources an “agentic-first” CRM after an $80k/year HubSpot quote</strong> — MIT licensed, and agentic-first means the core primitive is durable research agents running against accounts, not a form-and-pipeline app with an AI sidebar. The origin story is the payload: HubSpot quoted $80k a year for a sales team under ten people, so he spent half a day building a replacement. The clearest public data point yet for the build-vs-buy inversion — whether or not this particular CRM is good, the pricing pressure it represents will start showing up in vendor renewals (<a href="https://x.com/lewiscarhart/status/2083610805069611230">lewiscarhart</a>).</p><p><strong>Semantica: an open-source Palantir-style intelligence layer for agents</strong> — Three capabilities layered over agent operations: context graphs as a semantic model of entities and relationships, explainable reasoning where the chain from question to answer is inspectable, and a permanent audit trail of what the agent did and why. Deliberately the Foundry/AIP shape — the value isn’t the model, it’s the governed substrate around it. The enterprise objection to agents is almost never capability, it’s accountability: who saw what, what did it touch, can you reconstruct the decision for an auditor next year (<a href="https://x.com/tom_doerr/status/2083434480723718478">tom_doerr</a>).</p><p><strong>A local control plane for the rules, skills and hooks sprawled across your agents</strong> — Trevin Chow names the problem: rules, skills, hooks and accumulated corrections live separately in Claude Code, Codex and Cursor, and together form a dependency graph nobody can see or maintain. He points at blume.codes, building exactly that view locally. Modest reach but a bookmark-heavy ratio — the shape of a pain point people expect to return to. This is the config-drift failure mode in its pure form: patching an installed copy instead of the canonical upstream, or a wiped .env silently changing behaviour (<a href="https://x.com/trevin/status/2084679477842608223">trevin</a>).</p><p><strong>Have Claude map your app architecture into a self-explaining HTML + JSON pair</strong> — The HTML is the human-readable view; the JSON is machine-readable context for the next agent picking up the codebase, so the repo effectively explains itself. A lightweight pattern for persistent, agent-consumable architecture docs that survive across sessions and hand-offs, cutting the ramp-up cost every time an agent meets unfamiliar code (<a href="https://x.com/vivek4real_/status/2079625956176613758">Vivek4real_</a>).</p><p><strong>Onton claims a “successor architecture for search” with Ontology 1</strong> — An AI model aimed at ecommerce search, claimed at least 2.7× more accurate than the best existing engines, learning on its own with no retraining or fine-tuning, and hallucination-free. Read it sceptically: “hallucination-free” and “no retraining” are the two claims that most often do heavy definitional lifting, and the 2.7× is self-published against an unnamed baseline. But the underlying idea matters for catalog search where inventory changes constantly — a retrieval architecture that self-updates without a retraining loop would be a real unlock (<a href="https://x.com/onton_ai/status/2082511721856409939">onton_ai</a>).</p><p><strong>Google expands the Gemini family with three new models</strong> — Framed around production readiness — faster, more token-efficient, reliable at scale — rather than headline benchmark scores, which points at high-volume latency-sensitive deployments. No names or numbers in the announcement itself. The framing is the signal: frontier labs are now competing on cost-per-token and throughput as loudly as on capability (<a href="https://x.com/google/status/2079589747366724030">Google</a>).</p><h3>Engineering / DevOps</h3><p><strong>Perplexity open-sources Numbat, a harness-agnostic agent detection and response layer</strong> — Built to work across agent harnesses rather than bolted to one framework, it gives security teams visibility into what agents actually do on client endpoints plus controls to block selected actions <em>before</em> they execute — a pre-execution gate, not after-the-fact log review. Covers macOS, Linux and Windows for detection, prevention and investigation. As coding and desktop agents become standing installs on employee machines, this is the first credible open answer to the question every security team eventually asks (<a href="https://x.com/perplexity_ai/status/2082511900580196596">perplexity_ai</a>).</p><p><strong>OpenAI quietly ships an open-source Codex Security CLI — Hacker News found it first</strong> — Released without an announcement, then acknowledged on X. It scans repositories for security issues, tracks findings across runs so you see new-versus-known rather than a fresh wall of noise, verifies that a fix actually closed a finding, and plugs into CI/CD as a gate. The cross-run finding tracking is the part most static scanners get wrong, and it’s what makes this a credible LLM-native complement to existing scorecard checks (<a href="https://x.com/OpenAI/status/2082263717916586117">OpenAI</a>).</p><p><strong>Claude Security plugin for Claude Code enters beta</strong> — Scan your changes for vulnerabilities before committing, or run a full scan across an entire codebase, from the terminal, on the Claude inference you already pay for. Bringing security scanning into the local dev loop rather than CI or post-merge lowers the friction of catching things early. For any org standardising on Claude Code, it’s the obvious shift-left addition (<a href="https://x.com/claudeai/status/2079990597973057691">claudeai</a>).</p><p><strong>“quill” — a free, local, open-source replacement for Granola</strong> — Part of one developer’s declared “war on paid software that should be free.” A macOS menu-bar app: ⌘R to record, ⌘O to open the folder, full audio transcribed locally with no cloud round-trip and a deliberately minimal three-item menu — no accounts, no workspace, no upload. Heavy bookmark ratio. The honest limitation: it produces raw transcripts, not structured summaries, which is exactly the part a meeting-harvest pipeline depends on (<a href="https://x.com/dremnik/status/2081879207509217473">dremnik</a>).</p><p><strong>“parrot” — local Whisper dictation as a Wispr Flow replacement</strong> — The follow-up in the same series. Hold fn to dictate, whisper-base.en shown right in the menu bar, everything on-device, no subscription, no audio leaving the machine. Smaller reach than quill but the same signal, and it&#39;s the one worth internalising: local Whisper-class models are now good enough that a weekend-sized app credibly displaces a funded SaaS product for a single user. The moat moves to sync, team features and accuracy at the tail (<a href="https://x.com/dremnik/status/2082142748187267532">dremnik</a>).</p><p><strong>Hetzner quietly ships a free, no-SLA open-weight inference API</strong> — DeepSeek v4, GLM 5.2 and Kimi, on what Hetzner explicitly calls an exploratory platform: free, no SLAs, offered so people test their use cases and report what breaks, with no promise it becomes a permanent product. ~343K views, which is the real signal — a bare-metal host entering inference is being read as market news. A low-cost EU-based option matters for data residency and egress economics, but “free, no SLA, may be withdrawn” makes it a prototyping surface, not something to put behind a product path (<a href="https://x.com/maxjendrall/status/2087113184423333939">maxjendrall</a>).</p><p><strong>Three OSS projects that treat coding agents as infrastructure</strong> — OpenSRE is an AI SRE agent that takes an alert, pulls evidence from logs, metrics, configs and dependencies, tests hypotheses in parallel and posts a root-cause report back to Slack, with Datadog, Grafana, AWS and GitHub integrations. Opencove replaces split-pane agent UIs with an infinite canvas where Claude Code/Codex sessions, terminals, tasks and notes coexist and survive restarts. Vibeyard is an IDE for coding agents with multi-session projects, a grid view of every agent at once, live cost and context tracking, and P2P session sharing over WebRTC. The framing: the wave moved on from “look, AI can code” to infrastructure for running it (<a href="https://x.com/filicroval/status/2053094131690607034">filicroval</a>).</p><p><strong>OpenSRE — open-source AI SRE agents for your infrastructure</strong> — From Tracer-Cloud, public and self-hostable. Autonomous incident triage: ingest an alert, gather evidence across observability and config surfaces, reason toward a root cause instead of leaving an on-call engineer grepping dashboards. A maturing “agentic ops” category aimed squarely at MTTR and on-call toil (<a href="https://x.com/tom_doerr/status/2065657764936851484">tom_doerr</a>).</p><h3>Product Management</h3><p><strong>Gemini in Google Slides generates on-brand, fully editable decks from your own files</strong> — Available now: point it at your Docs, Sheets and PDFs and it builds a presentation in minutes, grounded in your source documents rather than a blank prompt. The differentiator is that every element stays 100% editable — it avoids the locked-artifact problem that makes most AI deck tools a dead end the moment you need to change one slide. Start at slides.new (<a href="https://x.com/googleworkspace/status/2079971445950632212">GoogleWorkspace</a>).</p><h3>Others</h3><p><strong>The AI spend gap is 625× — median $12/employee/month versus top 1% at $7,500</strong> — Olivia Moore at a16z, on Ramp card-spend data, with the note that she’s not sure we’ve ever seen an adoption gap quite like it. Two caveats before anyone quotes the figure: it’s card spend across Ramp’s customer base, not a market survey, and a per-employee average hides how concentrated the spend is inside those top firms — a handful of heavy agent workloads can carry the whole number. The interesting question isn’t whether anyone spends too much. It’s whether the gap correlates with output, which nobody in the thread demonstrates (<a href="https://x.com/omooretweets/status/2088314089659662799">omooretweets</a>).</p><p><strong>Stripe reportedly in talks to acquire OpenRouter for ~$10B</strong> — OpenRouter is the dominant unified API and routing layer in front of hundreds of models, so a Stripe deal would fold model-access billing and metering into the payments rails — a serious vertical move into AI infrastructure. Treat it as an unconfirmed rumour, but the price tag is the signal: $10B for what is essentially an aggregation and billing intermediary tells you how much value is accruing to the metering layer of the stack rather than the apps on top of it (<a href="https://x.com/polymarketmoney/status/2080466494832497004">PolymarketMoney</a>).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=924ec9755e09" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Context Window — July 2026]]></title>
            <link>https://medium.com/@smnbss/context-window-july-2026-3d2164846446?source=rss-5de037daf82f------2</link>
            <guid isPermaLink="false">https://medium.com/p/3d2164846446</guid>
            <category><![CDATA[ai-sharing]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Simone Basso]]></dc:creator>
            <pubDate>Mon, 20 Jul 2026 07:08:43 GMT</pubDate>
            <atom:updated>2026-07-20T07:08:43.691Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Context Window is monthly now.</strong> I ran it weekly for a quarter and the honest truth is a week is too short a lens — half the “news” was noise that didn’t survive to Friday. A month lets the signal separate from the spike. So this is the first monthly issue, and it covers everything since the last one: mid-June through mid-July. Five weeks of my timeline, distilled. Here’s what actually mattered.</p><p><strong>Cost stopped being a finance problem and quietly became an architecture decision.</strong> This was the loudest pattern of the whole stretch, and it came from operators, not vendors. Ramp’s engineering lead laid out why you can overspend and underuse AI at the same time, and why the fix is defaults, tiering and automation guardrails rather than budgets (<a href="https://x.com/rahulgs/status/2068046943276663154">rahulgs</a>). Coinbase reportedly halved its token bill without cutting token volume — just cheaper-model routing plus caching (<a href="https://x.com/gergelyorosz/status/2070735111226847242">GergelyOrosz</a>). Brian Armstrong turned it into five levers: open-weight defaults, request routing, cache optimisation, lean context, usage visibility (<a href="https://x.com/brian_armstrong/status/2070670644577280109">brian_armstrong</a>). DoorDash showed the endgame — an internal benchmark (“DashBench”) that let them blend a cheap open model into their code reviewer and come out <em>better and cheaper</em> (<a href="https://x.com/andyfang/status/2074252174226493584">andyfang</a>). Here’s the thing: none of these teams reached for a spend alert. They treated cost as a property of the system — routing, caching, model selection — not a line item to police. That’s exactly the multi-LLM, no-vendor-lock-in stance we keep defending. It’s nice to see the biggest engineering orgs arrive at the same place.</p><p><strong>Open weights caught up, and that turns “which model” into a question about sovereignty.</strong> The month was full of people running frontier-class models on their own hardware. GLM 5.2 got claimed at Opus-4.8 parity, running 100% local on a Mac Studio at 2-bit quant (<a href="https://x.com/alexfinn/status/2067751557379297762">AlexFinn</a>). Kimi K2.7’s trillion parameters were squeezed to 325GB and 40+ tokens/sec on a workstation (<a href="https://x.com/onusoz/status/2066508979644387468">onusoz</a>). Lisa Su ran a 235B model on a lunchbox-sized PC with no cloud (<a href="https://x.com/0xjokker/status/2066548789633323200">0xJokker</a>). Treat the “beats Opus” claims with the skepticism they deserve — they’re opinions, not benchmarks. But the <em>direction</em> is not debatable: capable models now run privately, cheaply, on kit you own, and nobody — no vendor, no government — can switch them off. The part nobody’s pricing in is the governance tail. An uncensored Qwen fine-tune hit ~3M downloads, reportedly never refusing a prompt (<a href="https://x.com/neural_avb/status/2073053073434722554">neural_avb</a>). Once weights are out, de-alignment and mass distribution are unstoppable. Sovereignty cuts both ways.</p><p><strong>“Loop engineering” grew up.</strong> Last issue the story was <em>the loop, not the prompt, is the unit of AI work.</em> This month it got a spine. What’s framed as Anthropic’s internal loop-engineering playbook surfaced — five principles (Discovery, Handoff, Verification, Persistence, Scheduling) and one rule that matters more than the rest: never let the generator grade its own work. Separate the generator from a skeptical evaluator (<a href="https://x.com/milesdeutscher/status/2071528356152303770">milesdeutscher</a>). LangChain shipped an entire Agent Development Platform built on the same premise and coined “loopcraft” (<a href="https://x.com/sydneyrunkle/status/2066928783534289358">sydneyrunkle</a>). Matthew Berman launched a Loop Library so you can reuse other people’s loops instead of reinventing them (<a href="https://x.com/matthewberman/status/2067682870009925736">MatthewBerman</a>). The prompt-engineering era optimised the perfect instruction. This era optimises the harness — generator/evaluator separation, verification gates, persistence. That’s structural. It’s the difference between an agent that occasionally impresses you and one you can leave running.</p><p><strong>AI climbed out of the chat box and into the team.</strong> For two years “using AI” meant a request/response window. That’s ending. Anthropic launched Claude Tag — Claude as a proactive Slack teammate with its own identity, memory and multiplayer collaboration, which Boris Cherny says changed how Anthropic works internally (<a href="https://x.com/bcherny/status/2069474681749754272">bcherny</a>). Slack turned Slackbot into an MCP host that fires 20+ apps in-thread (<a href="https://x.com/slackhq/status/2067638851938463962">SlackHQ</a>). Linear put external agents into a shared multiplayer instance where filing a bug can trigger a fix (<a href="https://x.com/karrisaarinen/status/2069859623138480389">karrisaarinen</a>). The most important detail is the least flashy one: when Claude operates in a shared channel it acts under <em>its own</em> credentials, not by borrowing a human’s (<a href="https://x.com/claudedevs/status/2069895377080443271">ClaudeDevs</a>). That’s the unlock. You can’t trust an agent with write actions in a team setting until permissions belong to the agent, are auditable, and follow least-privilege. Identity is the boring plumbing that makes the ambient-teammate story real — and it just got built.</p><p><strong>The company brain became the moat.</strong> The sharpest strategic post of the month wasn’t about a model at all. Zapier’s CEO is “killing the DM” internally, starting with the exec team, because one-to-one DMs lock company context in private silos no agent can reach (<a href="https://x.com/wadefoster/status/2072665341650407492">wadefoster</a>). Jacob Posel drew the ladder explicitly: a harness wired to your systems puts you top 1%, a way to share skills across it top 0.1%, and a company brain — shared knowledge, rules, learnings, sessions — makes you a unicorn (<a href="https://x.com/jacob_posel/status/2073883761113010555">jacob_posel</a>). Cerebras published how they actually built theirs (<a href="https://x.com/dbredvick/status/2078150905078206789">DBredvick</a>); Google proposed OKF as a portable “LLM-wiki” format (<a href="https://x.com/googlecloudtech/status/2067012903337664886">GoogleCloudTech</a>). Connectors are table stakes now. The durable advantage is queryable context plus shared skills plus a persistent brain. I’ll say the quiet part out loud: this is precisely the bet behind our own gbrain + skills + connectors stack, and watching the market converge on it is the best validation I could ask for.</p><p><strong>Agents got their own computer — and started running without you watching.</strong> Two shifts landed at once. Peter Steinberger’s line — to <em>really</em> end-to-end test software, give the agent its own computer, not a share of yours (<a href="https://x.com/steipete/status/2073214429655883814">steipete</a>) — captured the first: agents that operate a whole machine to verify their own work. levelsio streamed a live cloud iOS simulator built by Claude Code on a headless Mac Mini (<a href="https://x.com/levelsio/status/2075328941317886210">levelsio</a>); Claude Code shipped a built-in browser so it reads and clicks real sites (<a href="https://x.com/matthewberman/status/2075636186685775964">matthewberman</a>). The second shift is autonomy over time: Google’s Gemini API managed agents added background execution, scheduled triggers, budget caps and a free tier (<a href="https://x.com/googleaistudio/status/2077801843720093867">GoogleAIStudio</a>), and Gemini Spark became a 24/7 personal agent acting on smart triggers (<a href="https://x.com/geminiapp/status/2072395690345714159">geminiapp</a>). Put them together: the unit of work is becoming an always-on worker with its own environment and a spend limit. Not a chat you babysit.</p><p><strong>And the bottleneck isn’t the model anymore. It’s your org.</strong> Boris Cherny hears the same story everywhere: one person is 10x-ing their output with AI while the rest of the team lags, and he distilled the fix into four steps of adoption (<a href="https://x.com/bcherny/status/2077929379661844559">bcherny</a>). Sundar Pichai’s framing — learn to orchestrate agents now or spend 2027 catching up (<a href="https://x.com/zodchiii/status/2066905647841522077">zodchiii</a>). Kun Chen documented a workflow shipping 40+ PRs a day and it took 20 hours just to record (<a href="https://x.com/kunchenguid/status/2068367773533667565">kunchenguid</a>). Here’s the gut punch. Every theme above is about capability racing ahead — cheaper, more open, more autonomous, more embedded. But capability was never going to be the constraint. The constraint is how fast the <em>practice</em> spreads inside your company. In 2026 the gap between your best AI-native engineer and your median one is the single most important number on your org chart. Close it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*dqV8G1ziMnthK4XtXkKvZQ.png" /></figure><h3>AI / Agents</h3><p><strong>Ramp: cost discipline through defaults, not budgets</strong> — Ramp’s eng lead argues orgs overspend and underuse AI simultaneously. Their fix: lower company-wide reasoning defaults, track p50/p75/p95 session cost per PR/ticket and compress model tiers over time, and ban automations from frontier models. Directly relevant to how any org should govern LLM spend (<a href="https://x.com/rahulgs/status/2068046943276663154">rahulgs</a>).</p><p><strong>Coinbase halved its LLM bill via routing + caching</strong> — Roughly 50% lower dollar spend at the same token volume, by routing to performant-enough cheaper models (GLM 5.2, Kimi K2.7) and caching aggressively. The signal: model choice is a cost-optimisation surface, not a default-to-frontier decision (<a href="https://x.com/gergelyorosz/status/2070735111226847242">GergelyOrosz</a>).</p><p><strong>Brian Armstrong’s five levers for flat AI spend</strong> — Hold costs flat while token use grows exponentially through engineering defaults, not spend alerts: open-weight defaults, intelligent routing, cache optimisation, lean context management, usage visibility. Treat cost as an architectural property (<a href="https://x.com/brian_armstrong/status/2070670644577280109">brian_armstrong</a>).</p><p><strong>DoorDash blends open-weight models using its own benchmark</strong> — “DashBench” lets DoorDash confidently mix a cheap open model (Kimi K2.6) into their AI code reviewer for the easy work and reserve Fable for the hard calls — coming out better <em>and</em> cheaper. The lesson: owning a domain-specific benchmark is what unlocks tiering (<a href="https://x.com/andyfang/status/2074252174226493584">andyfang</a>).</p><p><strong>Coding subscriptions are far more generous than API math implies</strong> — SemiAnalysis bought every Anthropic/OpenAI plan and ran tasks to the weekly limit, finding the effective token allowance well exceeds the naive “$200 ≈ $2K of API tokens” heuristic. Labs are subsidising power users to lock in subscriptions (<a href="https://x.com/semianalysis_/status/2064815044085318040">semianalysis_</a>).</p><p><strong>GLM 5.2 claimed at Opus-4.8 parity, running 100% local</strong> — Alex Finn reversed his “open trails frontier by 6 months” stance, running GLM 5.2 at 2-bit on a Mac Studio (~250GB) for a 24/7 coding loop. Treat the parity claim as opinion, not benchmark — but the local-sovereignty direction is real (<a href="https://x.com/alexfinn/status/2067751557379297762">AlexFinn</a>).</p><p><strong>Kimi K2.7’s 1T parameters shrunk to 325GB</strong> — Unsloth’s dynamic 2-bit quant cut the trillion-parameter model by ~48% to 325GB, running 40+ tok/s on high-end workstations. Trillion-parameter models are reaching local deployability (<a href="https://x.com/onusoz/status/2066508979644387468">onusoz</a>).</p><p><strong>A 235B model on a lunchbox-sized PC</strong> — Lisa Su demoed a 235B model running live on an AMD Ryzen AI Max+ 395 mini-PC — no data centre, no cloud. A proof point that near-consumer hardware can handle frontier-class inference (<a href="https://x.com/0xjokker/status/2066548789633323200">0xJokker</a>).</p><p><strong>Magnitude: an open-model coding agent, ~60% cheaper than Claude Code</strong> — Runs entirely on open weights, `npm i -g @magnitudedev/cli`, claiming no performance drop. Another data point that open models are becoming credible substitutes in agentic coding (<a href="https://x.com/tomgreenwald/status/2068008885018062868">tomgreenwald</a>).</p><p><strong>An uncensored Qwen3.6 fine-tune hit ~3M downloads</strong> — A de-aligned variant that reportedly never refuses a prompt, spreading fast on Hugging Face. The reminder that open-weight releases carry a governance tail: once weights are out, downstream de-alignment is unstoppable (<a href="https://x.com/neural_avb/status/2073053073434722554">neural_avb</a>).</p><p><strong>Anthropic’s “loop engineering” playbook — 5 principles + generator/evaluator</strong> — Discovery, Handoff, Verification, Persistence, Scheduling, plus a hard rule: never let the generator grade its own work — use a separate skeptical evaluator. Six architectural components and a “barbell” cost approach (cheap models for bulk, expensive for the hard calls) (<a href="https://x.com/milesdeutscher/status/2071528356152303770">milesdeutscher</a>).</p><p><strong>LangChain ships an Agent Development Platform built on “loopcraft”</strong> — LangSmith Deployment + Engine + a deepagents quickstart, positioning reliable agents as a matter of robust execution loops, not individual prompts. Opinionated architecture guidance shipped alongside infrastructure (<a href="https://x.com/sydneyrunkle/status/2066928783534289358">sydneyrunkle</a>).</p><p><strong>Loop Library: a reusable directory of agent loops</strong> — A community-submittable catalogue of repeatable autonomous workflows you can drop into your own setup. Agent behaviours are becoming shareable primitives rather than one-off prompts (<a href="https://x.com/matthewberman/status/2067682870009925736">MatthewBerman</a>).</p><p><strong>Claude Tag: Claude as a proactive Slack teammate</strong> — Its own identity, persistent memory, multiplayer collaboration, and scoped channel/tool access. Cherny says it “totally changed” how Anthropic uses Claude internally — a direct shot at the “AI coworker” category (<a href="https://x.com/bcherny/status/2069474681749754272">bcherny</a>).</p><p><strong>Claude acts under its own credentials in shared channels</strong> — Rather than borrowing a human participant’s permissions, the agent has its own identity — cleanly separating what it may do from who’s in the conversation. The precondition for trusting agents with write actions and auditability (<a href="https://x.com/claudedevs/status/2069895377080443271">ClaudeDevs</a>).</p><p><strong>Claude runs incident response end-to-end from a thread tag</strong> — Tag Claude when the page fires: it pulls graphs, diffs the offending deploy, proposes a root cause; the team approves; it opens the PR, lands it, watches the metric recover, and resolves the page. Agents closing the loop on on-call work, not just summarising alerts (<a href="https://x.com/claudedevs/status/2069468902216945939">ClaudeDevs</a>).</p><p><strong>Linear puts external agents in a shared multiplayer instance</strong> — Agents write code, automate triage and bug-fixing as first-class participants alongside humans; the CEO demoed going from bug report to PR on mobile. The issue tracker becomes an execution surface (<a href="https://x.com/karrisaarinen/status/2069859623138480389">karrisaarinen</a>).</p><p><strong>Slackbot becomes an MCP host</strong> — An MCP client built into Slackbot invokes 20+ partner apps in-thread — build with Replit, pull a live Amplitude chart, update a Linear ticket, make a Canva design — turning chat into an action surface (<a href="https://x.com/slackhq/status/2067638851938463962">SlackHQ</a>).</p><p><strong>Open Tag: an open-source, model-agnostic Claude Tag</strong> — Any model, any harness, custom agents, with generative UI, streaming, and human-in-the-loop approvals, launching on Slack and Teams. A portable way to embed agentic chat experiences without binding to one vendor (<a href="https://x.com/ataiiam/status/2070269772126937456">ataiiam</a>).</p><p><strong>Zapier is “killing the DM” to make context queryable</strong> — Private DMs lock company context in silos no agent can reach; pushing communication into open channels makes org context a substrate for AI. AI-readiness is information architecture, not just tooling (<a href="https://x.com/wadefoster/status/2072665341650407492">wadefoster</a>).</p><p><strong>The AI-native ladder: harness → shared skills → company brain</strong> — Top 1% have a harness wired to all core systems; top 0.1% add a system for sharing skills; “unicorns” have a company brain for shared knowledge, rules, learnings and sessions. A clean external benchmark for the gbrain-style bet (<a href="https://x.com/jacob_posel/status/2073883761113010555">jacob_posel</a>).</p><p><strong>Cerebras publishes how it built its company knowledge base</strong> — A concrete, public reference architecture for an internal “company brain” — the retrieval-and-reasoning layer that turns scattered docs into a queryable system. Rare and directly useful prior art (<a href="https://x.com/dbredvick/status/2078150905078206789">DBredvick</a>).</p><p><strong>Google proposes Open Knowledge Format (OKF)</strong> — An open spec formalising the “LLM-wiki” pattern into a portable, interoperable standard, with a BigQuery enrichment agent and an HTML visualiser. Could become the OpenAPI of org knowledge ↔ agents (<a href="https://x.com/googlecloudtech/status/2067012903337664886">GoogleCloudTech</a>).</p><p><strong>Give the agent its own computer for true end-to-end testing</strong> — Steinberger’s argument: isolation lets an agent drive the full stack — real UI, real services — without contaminating your dev box. A shift from agents that edit code to agents that operate a machine to verify their own work (<a href="https://x.com/steipete/status/2073214429655883814">steipete</a>).</p><p><strong>Gemini API managed agents: background, triggers, budgets, free tier</strong> — Server-hosted agents get long-running background execution (poll an ID, not a fragile HTTP connection), remote MCP, mid-session credential refresh, scheduled triggers, budget caps, and a free tier. Managed agents become asynchronous, cron-like workers (<a href="https://x.com/googleaistudio/status/2077801843720093867">GoogleAIStudio</a>).</p><p><strong>Gemini Spark becomes a 24/7 personal agent</strong> — A native macOS app plus “smart triggers” that act on events automatically — e.g. auto-emailing a game analysis after your team plays. Consumer “ambient agents” going mainstream, gated to the AI Ultra tier (<a href="https://x.com/geminiapp/status/2072395690345714159">geminiapp</a>).</p><p><strong>A SKILL.md markdown file as an agent’s router</strong> — Google’s `gemini-anything-agent` on managed agents classifies intent and reaches for the narrowest tool itself (image, video, music, TTS, web app); the “router” is a versionable `SKILL.md` + tiny CLI discovered via ` — help`. Any CLI you can describe becomes an agent capability — a far better contract than a model-picker dropdown (<a href="https://x.com/lyalindotcom/status/2073128677055819785">lyalindotcom</a>).</p><p><strong>Claude Code artifacts can call MCP connectors at runtime</strong> — Artifacts fetch live data and act on behalf of each viewer instead of being static snapshots — per-viewer dashboards and mini-apps — on paid plans, but not on publicly-shared artifacts (<a href="https://x.com/claudedevs/status/2077489907350856038">claudedevs</a>).</p><p><strong>Codex goes model-agnostic</strong> — Codex CLI officially runs non-OpenAI and local models via Ollama/LM Studio or third-party APIs, turning it into a model-agnostic coding harness you keep while swapping the backing model for cost, privacy or sovereignty (<a href="https://x.com/hqmank/status/2067245758202728771">hqmank</a>).</p><p><strong>Google’s Agentic Resource Discovery (ARD) open spec</strong> — A decentralized way for agents to discover and <em>trust</em> each other’s tools, skills, MCP servers and capabilities across the web, backed by a broad coalition (GitHub, NVIDIA, Databricks, Microsoft, Salesforce, Snowflake…). The discovery/trust layer next to MCP and A2A (<a href="https://x.com/googledevs/status/2067291188919468510">googledevs</a>).</p><p><strong>Sakana AI’s “Fugu”: multi-agent orchestration behind one API</strong> — Frontier-class capability collapsed into a single model endpoint rather than a framework you wire yourself, pitched from Japan as “export-control-free.” “Multi-agent system as a single model” is becoming a product category (<a href="https://x.com/sakanaailabs/status/2068861630327443966">SakanaAILabs</a>).</p><p><strong>SpaceX acquires Cursor in an all-stock deal</strong> — Cursor folds into the SpaceX/xAI umbrella to build “the world’s best coding and knowledge-work AI” — potentially the year’s most consequential dev-tools consolidation (<a href="https://x.com/spacex/status/2066873915717136548">SpaceX</a>).</p><p><strong>“Origin”: a Git built for agent workloads</strong> — From Cursor/Graphite, a version-control system designed for the parallel, non-linear commits agents generate, with API/MCP extensibility and an auto-resolution agent. Traditional Git breaks down under agent volume (<a href="https://x.com/swyx/status/2066928345246470204">swyx</a>).</p><h3>Engineering / DevOps</h3><p><strong>Cloudflare’s Monetization Gateway settles in stablecoins via x402</strong> — Charge for any page, dataset, API or MCP tool behind Cloudflare, settled over the HTTP-native x402 protocol (reviving the 402 “Payment Required” code). A built-in paywall autonomous agents can pay programmatically — Cloudflare at the toll booth for the agentic web (<a href="https://x.com/cloudflare/status/2072311802285723953">cloudflare</a>).</p><p><strong>OpenShip: an open-source, self-hosted PaaS</strong> — Git-based zero-downtime deploys with one-click rollbacks, per-branch environments, one-click managed services (Postgres, Redis, Kafka, ClickHouse…), plus MCP support so agents can manage infrastructure directly. A Coolify-style escape from hyperscaler cost and lock-in, agent-native from day one (<a href="https://x.com/engmoelgaraihy/status/2078456995284148409">EngMoElgaraihy</a>).</p><p><strong>GPT-5.5-Cyber + an updated Codex Security plugin</strong> — OpenAI’s most capable cyber model, purpose-built for <em>authorized defensive</em> work — finding vulnerable code, validating issues, developing patches — wired into the coding workflow. A credible step toward continuous, agent-assisted vulnerability remediation (<a href="https://x.com/clintgibler/status/2069487614982541765">clintgibler</a>).</p><p><strong>CodexBar now brokers 56 AI providers</strong> — A menu-bar utility for AI coding providers gets a redesign and now spans 56 providers via community PRs. A telling data point on how fragmented the coding-tool landscape has become — you need an aggregation layer (<a href="https://x.com/steipete/status/2073127743131464059">steipete</a>).</p><h3>Design</h3><p><strong>Claude Design ships WYSIWYG editing + two-way Claude Code integration</strong> — Real WYSIWYG editing, design-systems import, and — the standout — two-way sync with Claude Code so designs and code stay aligned instead of a one-shot handoff. AI design tooling moving toward a continuous designer-developer loop (<a href="https://x.com/nateparrott/status/2067333017429766481">nateparrott</a>).</p><p><strong>Auto-mapping every UI state to fix the “design source of truth” problem</strong> — With coding agents 10x-ing change throughput, manual Figma-vs-implementation review is impossible; a mobile-use agent (Revyl) mapped every screen state of an app in under an hour. An always-current map lets humans and agents ship without visual blindspots (<a href="https://x.com/anamhira/status/2069811537091457221">anamhira</a>).</p><p><strong>`/visual-plan`: wireframe whole user flows as storyboards</strong> — An open-source coding-agent skill reads your codebase and renders any flow as a storyboard so you spot UX problems visually — best aimed at the least-inspected, highest-leverage sign-up/onboarding paths (<a href="https://x.com/steve8708/status/2070162604434280917">Steve8708</a>).</p><p><strong>The AI-native design process, in six steps</strong> — Record planning meetings for context, feed inspiration images, generate dozens of live prototypes, spin up one-off mini-tools, maintain a `design.md`, ship. Prototyping cost collapses toward zero, so the designer’s job shifts to curation and combination (<a href="https://x.com/aaron_epstein/status/2075635266585297042">aaron_epstein</a>).</p><h3>Product Management</h3><p><strong>“Loop Engineering” as the next core PM skill</strong> — The argument that the defining PM skill after prompt engineering is designing feedback loops, evals and iteration as the unit of AI-product design. As model capability commoditises, durable advantage comes from the loop you build around the model, not the prompt you send it (<a href="https://x.com/saboo_shubham_/status/2068730090457006588">Saboo_Shubham_</a>).</p><h3>Others</h3><p><strong>A decade of angel investing returned just 7.8% IRR</strong> — A candid post-mortem: ~$300K across 53 deals via top AngelList syndicates (2014–2019) for a 7.8% IRR. A reality check that premium deal flow doesn’t guarantee venture-beating returns, and diversified angel portfolios can badly lag public markets (<a href="https://x.com/mediaking/status/2075022188332122501">MediaKing</a>).</p><p><strong>Twelve things to change rebuilding ElevenLabs’ GTM</strong> — A widely-shared thread on hiring, sales enablement, market expansion, outbound focus, and prioritising deal quality over volume. A useful reference for structuring a modern sales org at a fast-scaling AI company (<a href="https://x.com/carles_reina/status/2077075493946597749">carles_reina</a>).</p><p><strong>Builder.io’s “Clips”: an open-source, agent-native Loom</strong> — Free and forkable; every clip ships APIs and metadata so an agent can “see and hear” the video at any timestamp from its URL. A concrete example of the “feedback artifacts must be machine-readable” shift (<a href="https://x.com/steve8708/status/2069064659613958348">Steve8708</a>).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3d2164846446" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Context Window — W24 2026]]></title>
            <link>https://medium.com/@smnbss/context-window-w24-2026-aa004d131c22?source=rss-5de037daf82f------2</link>
            <guid isPermaLink="false">https://medium.com/p/aa004d131c22</guid>
            <category><![CDATA[ai-sharing]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Simone Basso]]></dc:creator>
            <pubDate>Wed, 17 Jun 2026 06:19:16 GMT</pubDate>
            <atom:updated>2026-06-17T06:19:16.986Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>The loop, not the prompt, is now the unit of AI engineering.</strong> Addy Osmani published a deep dive on “loop engineering” — building systems that recursively prompt their own agents toward a goal instead of you re-prompting at every step (<a href="https://x.com/addyosmani/status/2064127981161959567">addyosmani</a>). The same week, Rahul’s explainer made the identical argument to a much wider room, citing practitioners like Peter Steinberger and Boris Cherny moving from single-shot prompting to self-correcting loops where the agent discovers, plans, executes, verifies, and iterates on its own (<a href="https://x.com/sairahul1/status/2064277888216555684">sairahul1</a>). And Cathryn Lavery showed what that actually looks like on hardware: always-on AI “employees” running across a fleet of Macs, wired together with SSH, Tailscale, and reusable skills, a Mac mini agent stood up in 20 minutes (<a href="https://x.com/cathrynlavery/status/2064752541078901201">cathrynlavery</a>). Put them together and the shift is obvious. Prompt engineering was about the perfect single instruction. Loop engineering is about the system that keeps going without you. That’s the same bet behind our own agents/ registry and the super skills — the work is designing the loop, not babysitting it.</p><p><strong>AI traffic just became a number you can put on a dashboard.</strong> Google Analytics is rolling out automated measurement of visits arriving from AI assistants — the chatbots that have been quietly eating the top of the funnel (<a href="https://x.com/googleanalytics/status/2064742718664818957">googleanalytics</a>). No manual UTM tagging, no referrer guesswork: human traffic from the major assistants becomes a first-class, trendable channel inside GA. This matters because for the last year “AI search visibility” has been a black box — you knew clicks were leaking somewhere, you just couldn’t see where. Now you can. For SEO and growth teams that means AI-referred traffic stops being a vibe and starts being a KPI you can defend in a planning meeting. The measurement layer is finally catching up to the behaviour change. About time.</p><p><strong>A government can now switch off the model you built on.</strong> The week’s loudest post was Anthropic suspending access to Claude Fable 5 and Mythos 5 in response to a US export-control directive (<a href="https://x.com/claudedevs/status/2065597942602531163">ClaudeDevs</a>). Treat the specifics with caution — this is one viral post and the framing deserves verification against Anthropic’s own channels. But the scenario is the part worth sitting with. We spend a lot of energy arguing model quality and price per token, and almost none on the question this raises: what happens to your product when the model line you depend on becomes a geopolitical variable someone else controls? Availability is a dependency, same as any vendor. The answer isn’t panic. It’s the same boring discipline that’s always paid off — abstract the model, keep a fallback, don’t hard-wire your roadmap to one provider you can’t switch off yourself.</p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*dqV8G1ziMnthK4XtXkKvZQ.png" /></figure><h3>AI / Agents</h3><p><strong>Loop engineering replaces prompt engineering</strong> — Addy Osmani’s deep dive argues the new craft is building systems that recursively prompt their own agents toward a goal — discover, plan, execute, verify, iterate — instead of a human re-prompting at every step. It’s anchored by a sprawling set of companion essays on adversarial code review, agent harness design, the “factory model,” the “orchestration tax,” and new failure modes like intent debt, comprehension debt, and cognitive surrender. The framing positions loop design as the direct successor to prompt engineering. For anyone building agents, the takeaway is that the reusable asset is the loop itself — the scaffolding that keeps an agent productive unattended — not the perfectly worded single instruction (<a href="https://x.com/addyosmani/status/2064127981161959567">addyosmani</a>).</p><p><strong>“Loops: what every AI engineer needs to know in 2026”</strong> — Rahul’s widely-read explainer makes the loop argument to a mainstream audience: senior AI engineers have moved from prompting agents once to designing autonomous feedback cycles where the agent self-corrects without constant supervision. It name-checks practitioners like Peter Steinberger and Boris Cherny as examples of the shift, and stresses economics — cheaper models such as DeepSeek make long-running loops financially viable, not just technically possible. Read next to Osmani’s piece, the two converge on the same thesis from different altitudes: the unit of AI engineering work is now the loop, not the prompt. The cost angle is the part product leaders should internalize — sustained autonomy is becoming affordable, which changes what a small team can attempt (<a href="https://x.com/sairahul1/status/2064277888216555684">sairahul1</a>).</p><p><strong>Always-on AI “employees” running across a fleet of Macs</strong> — Cathryn Lavery published a hands-on guide to running persistent AI agents across multiple Macs, stitched into one workspace with SSH, Tailscale, and reusable agent skills, with agents (OpenClaw / Hermes) reachable from anywhere and a Mac mini “AI employee” stood up in about 20 minutes. It’s the concrete counterpart to the conceptual loop-engineering threads — what the always-on agent fleet actually looks like on personal hardware rather than in the cloud. The signal for tech leaders: self-hosted, always-running agent infrastructure is now a weekend project, which pulls “give every person a team of agents” much closer to reach than the org-chart conversations suggest (<a href="https://x.com/cathrynlavery/status/2064752541078901201">cathrynlavery</a>).</p><p><strong>Anthropic reportedly suspends Claude Fable 5 and Mythos 5 over a US export directive</strong> — A heavily-shared post quotes an Anthropic statement saying the US government issued an export-control directive affecting access to both model lines, prompting a suspension. Treat it as unverified — it’s a single viral claim and the framing warrants checking against Anthropic’s own channels. But the scenario is the real lesson regardless of this specific case: model availability is a vendor dependency that can move for reasons entirely outside your control, including geopolitics. Any team that has hard-wired a product to one model line should read this as a prompt to abstract the model and keep a working fallback, not to react to the headline itself (<a href="https://x.com/claudedevs/status/2065597942602531163">ClaudeDevs</a>).</p><h3>SEO</h3><p><strong>Google Analytics adds automated AI-assistant traffic measurement</strong> — GA is rolling out built-in tracking of human visits that arrive from AI assistants, which Google frames as closing a “massive blind spot” as AI answers shift where clicks originate. It surfaces and trends traffic from the major chatbots directly in GA reports — no manual UTM or referrer tagging required — turning AI-referred visits into a first-class, trendable channel. For SEO and growth teams this is an early sign that “AI search visibility” is becoming a measurable KPI alongside classic organic traffic, rather than an untracked unknown you can only infer from declining clicks. The practical move is to start trending this channel now, before AI cannibalization of organic becomes a budget conversation (<a href="https://x.com/googleanalytics/status/2064742718664818957">googleanalytics</a>).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=aa004d131c22" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Context Window — W23 2026]]></title>
            <link>https://medium.com/@smnbss/context-window-w23-2026-47c2ae69b44e?source=rss-5de037daf82f------2</link>
            <guid isPermaLink="false">https://medium.com/p/47c2ae69b44e</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai-sharing]]></category>
            <dc:creator><![CDATA[Simone Basso]]></dc:creator>
            <pubDate>Wed, 10 Jun 2026 18:11:48 GMT</pubDate>
            <atom:updated>2026-06-10T18:11:48.450Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>The harness is the moat now — and the whole industry said it out loud this week.</strong> Garry Tan called it directly: the <em>AI Harness Wars of 2027</em>, where labs stop competing on raw model IQ and start competing on the scaffolding that wraps the model into something useful (<a href="https://x.com/garrytan/status/2061878212213572083">garrytan</a>). The trigger was Factory Router auto-picking the best model per task and claiming frontier performance at 25% lower cost. Same week, Anthropic shipped ant, an official CLI that exposes every Claude Platform endpoint to the terminal so an agent can drive the platform exactly like a human does (<a href="https://x.com/minchoi/status/2061938760678555702">minchoi</a>), and a Claude Code engineer&#39;s &quot;harness engineering&quot; masterclass went mini-viral because operator-level craft is finally being treated as a discipline, not a prompt-tips thread (<a href="https://x.com/neural_avb/status/2061918166566195329">neural_avb</a>). Even the unit of teachable behaviour converged: Gemini Enterprise launched &quot;Skills&quot; — same name, same shape as Anthropic&#39;s — so both closed-model vendors now standardize on the Skill as the reusable building block (<a href="https://x.com/googlecloud/status/2062273231692726565">googlecloud</a>). Here&#39;s the thing: model capability per dollar is flattening, and everyone now agrees the differentiation moved up the stack. That&#39;s exactly the bet we&#39;re making with our own agents/ registry and the super skills. Routing, tooling, memory, evals. The model is the commodity. The harness is the asset.</p><p><strong>Coding-agent productivity finally has real numbers — and they’re the bookends of the whole debate.</strong> Uber reportedly capped coding-agent spend at $1,500 per employee, per month, per tool — Simon Willison called it sensible, but the louder signal is the implicit admission that a single engineer can plausibly burn that much, and that uncapped agent budgets are a live risk, not a theoretical one (<a href="https://x.com/simonw/status/2062143151184465964">simonw</a>). On the other end, Anthropic disclosed that its own engineers now ship roughly 8x as much code per quarter as their 2021–2025 baseline, framing it as a path to recursive self-improvement (<a href="https://x.com/anthropicai/status/2062568864240836995">AnthropicAI</a>). Hold those two next to each other: $1,500/seat of spend, 8x of output. That’s the actual ROI conversation every CTO is now having, including me. Stop arguing whether the uplift is real — it’s real enough to budget for and big enough to govern. The only open question is whether your team is the one driving the spend or the one capping it.</p><p><strong>Enterprise RAG grew up: verify before you answer.</strong> Google shipped an Agentic RAG framework on Gemini Enterprise that replaces “search once, grab a few docs, generate” with a coordinated set of agents — query optimization, retrieval, and a “Sufficient Context Agent” that checks whether the evidence is actually enough <em>before</em> the model is allowed to answer. They report up to 34% higher accuracy and 90.1% on cross-corpus retrieval (<a href="https://x.com/ai_for_success/status/2062984524968218690">ai_for_success</a>). It rhymes with Anthropic’s internal self-service analytics playbook from the same week — reusable Skills, proper data foundations, and rigorous evals before an agent touches production data (<a href="https://x.com/claudedevs/status/2062274312363770064">ClaudeDevs</a>). The naïve retrieve-and-generate loop is what hallucinates the moment your data spans Drive + Confluence + tickets — which is exactly our world. The sufficiency check is the pattern to steal for the Cass/Vertex work: make the system admit “I don’t have enough” instead of confidently making something up.</p><p><strong>AI search became a measurable, separate discovery layer — and the playbook is nothing like classic SEO.</strong> Ahrefs dropped a billion-datapoint study and the findings reorder the priorities: “best X” listicles are 43.8% of all ChatGPT citations, 67% of its top citations come from sources marketers can’t influence, 28.3% of its most-cited pages have <em>zero</em> Google organic visibility, and schema markup had no meaningful impact at all. The single highest correlation with AI brand visibility wasn’t backlinks or DR — it was YouTube mentions (0.737) (<a href="https://x.com/timsoulo/status/2061796432534003866">timsoulo</a>). Days later Google confirmed it’s finally shipping a dedicated AI performance report in Search Console — impressions only at first, no queries — giving us first-party visibility into AI surfaces that have been a measurement black hole until now (<a href="https://x.com/brodieseo/status/2062042288864244086">brodieseo</a>). For WeRoad this settles two things: YouTube is the under-priced AI-SEO channel, and the job is to <em>be the source of truth</em>, not to tweak schema. Flag GSC opt-in to the SEO team and start treating impressions-without-clicks as the leading indicator of AI cannibalization.</p><p><strong>The agent stack is moving off cloud APIs and onto local silicon — and the open-weight crowd is getting first-class treatment.</strong> The week’s biggest signal by a mile was Nous Research’s Hermes Agent running natively on NVIDIA’s new RTX Spark superchip, featured in the Computex keynote, plus an integration with Microsoft’s OpenShell runtime that wires Hermes into OS-level security primitives (<a href="https://x.com/nousresearch/status/2061323987804713083">NousResearch</a>). The interest was enormous — outsized even for Nous — and that’s the tell. An open-weight agent framework just got first-class silicon from NVIDIA and first-class OS plumbing from Microsoft in the same breath. Pair that with Google pushing multi-day, unattended autonomous agents managed from a unified “Inbox” (<a href="https://x.com/googlecloud/status/2061105927692919271">googlecloud</a>), and the trajectory is clear: agentic workloads are sliding down from cloud API calls toward edge hardware and longer-running autonomy. The cloud isn’t the only place agents will live for much longer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*dqV8G1ziMnthK4XtXkKvZQ.png" /></figure><h3>AI / Agents</h3><p><strong>Garry Tan calls the “AI Harness Wars of 2027”</strong> — The YC president argues that as raw model capability flattens and commoditizes, frontier labs will try to make their agent <em>harness</em> the moat — tooling, routing, memory, the scaffolding around the model. The trigger was Factory launching Factory Router, which auto-selects the best model per task while claiming to hold frontier performance at 25% lower cost. The take that matters: consumers win when model quality converges and differentiation moves up the stack. This is the architecture pattern to evaluate internally as Claude, GPT, and Gemini keep converging in capability per dollar (<a href="https://x.com/garrytan/status/2061878212213572083">garrytan</a>)</p><p><strong>Anthropic ships </strong><strong>ant, a CLI for the Claude Platform</strong> — ant exposes every API endpoint — Messages, Managed Agents, file uploads, YAML sync, run inspection — directly from the terminal, designed so coding agents can drive it. Anything a human runs interactively on the Claude Platform, an agent can now run too, piping results straight into shell pipelines. It collapses the gap between human and agent operators and is a clean building block for the agent-harness pattern we lean on across the super skills and the claude-api workflow (<a href="https://x.com/minchoi/status/2061938760678555702">minchoi</a>)</p><p><strong>Anthropic: engineers ship 8x more code per quarter than the 2021–2025 baseline</strong> — The most concrete public productivity number any frontier lab has shared about itself, dropped inside a thread arguing Claude is accelerating Anthropic’s own development — a “possible path to recursive self-improvement.” Whatever you make of the framing, 8x is the figure every CTO will now be held against when setting expectations for coding-agent uplift. Read it next to Uber’s $1,500/seat cap as the bookends of how much agent productivity is actually real (<a href="https://x.com/anthropicai/status/2062568864240836995">AnthropicAI</a>)</p><p><strong>Uber caps coding-agent spend at $1,500/employee/month/tool</strong> — Simon Willison flags Uber imposing a per-employee, per-tool monthly cap on agent usage. He calls it sensible, but the real signal is implicit: Uber clearly believes an individual engineer can drive that much spend, meaning the productivity value — and the burn rate — of Claude Code, Cursor, and Codex is materially higher than headline subscription pricing implies. A useful benchmark when WeRoad sets its own per-seat caps (<a href="https://x.com/simonw/status/2062143151184465964">simonw</a>)</p><p><strong>Gemini Enterprise adds “Skills” — reusable saved workflows</strong> — Google Cloud lets a user codify a recurring workflow (brand guidelines, report formatting) and save it as a reusable action the whole team can invoke. It’s the Anthropic Skills concept — same name, same shape — landing inside Gemini Enterprise. The interesting bit is the convergence: both major closed-model vendors now standardize on “Skill” as the unit of teachable agent behaviour, which has direct implications for how we structure our own agent registry (<a href="https://x.com/googlecloud/status/2062273231692726565">googlecloud</a>)</p><p><strong>Google ships Agentic RAG on Gemini Enterprise — 34% accuracy lift</strong> — Replaces classic single-shot RAG with coordinated agents for query optimization, retrieval, and a “Sufficient Context Agent” that verifies the retrieved evidence is enough before the model answers. Google reports up to 34% higher accuracy and 90.1% on cross-corpus retrieval. The sufficiency check directly targets premature answers and hallucinations — exactly what bites enterprise RAG once data spans Drive + Confluence + ticketing. A reference architecture for the Cass/Vertex work; the sufficiency-agent pattern is worth bolting onto our own retrieval (<a href="https://x.com/ai_for_success/status/2062984524968218690">ai_for_success</a>)</p><p><strong>Anthropic publishes its internal self-service analytics playbook</strong> — A long-form post on how Anthropic automated business analytics with Claude agents, across three pillars: reusable Skills, getting data foundations right (schemas, semantic layers, governance), and shipping rigorous evals before agents touch production data. It’s effectively a reference architecture for replacing ad-hoc analyst requests with self-service agents employees can trust — directly applicable to anything we build on top of BigQuery (<a href="https://x.com/claudedevs/status/2062274312363770064">ClaudeDevs</a>)</p><p><strong>ChatGPT rolls out cross-conversation memory</strong> — OpenAI upgraded ChatGPT to carry context reliably across conversations and keep it useful over time, rather than relying on one growing scratchpad that goes stale. The linked “memory dreaming” research frames it as new techniques for selecting, refreshing, and consolidating what the model remembers between sessions. A direct competitive datapoint for any Brainiac / Outline / Cass roadmap built on personal memory — the bar for “remembers me” just moved (<a href="https://x.com/openai/status/2062567556524003631">OpenAI</a>)</p><p><strong>Hermes Agent lands on NVIDIA RTX Spark with Microsoft OpenShell</strong> — Nous Research partnered with NVIDIA to run Hermes natively on the new RTX Spark superchip — featured in the Computex keynote — plus an OpenShell integration wiring Hermes into Microsoft’s security primitives. The outsized attention is the story: an open-weight agent framework getting first-class treatment from both NVIDIA silicon and Microsoft’s OS-level plumbing, a meaningful step toward agentic workloads moving off cloud APIs onto edge superchips (<a href="https://x.com/nousresearch/status/2061323987804713083">NousResearch</a>)</p><p><strong>Gemini Enterprise adds long-running autonomous agents with a unified Inbox</strong> — Google Cloud lets teams build and deploy agents that independently work complex problems for up to days at a time, managed through a new “Inbox” command center positioned as a way to scale an “agentic task force.” The headline differentiator is duration — multi-day, unattended execution. It’s a positioning signal that the enterprise agent battle is shifting from “can it call tools” to “can it run unattended for days and stay governable” (<a href="https://x.com/googlecloud/status/2061105927692919271">googlecloud</a>)</p><p><strong>A free “harness engineering” masterclass from the Claude Code team</strong> — AVB flagged a masterclass from a Claude Code engineer on harness engineering: the discipline of building the tools, prompts, memory, and retry/eval loops that wrap a frontier model into a useful agent. It struck a nerve fast because there’s real demand for serious craft material in this space rather than another prompt-tips thread — the kind of operator-level content normally kept as tribal knowledge inside labs (<a href="https://x.com/neural_avb/status/2061918166566195329">neural_avb</a>)</p><h3>Product Management</h3><p><strong>Linear opens its Team Docs beta to everyone after 200+ companies sign up in 24h</strong> — One of Linear’s most-requested features, Team Docs moves the issue tracker into team-documentation territory, with Team Docs and Home as the new connective tissue between specs, issues, and async comms. Notably, Karri Saarinen used Linear Agent to mine 40,000+ customer requests to find the pattern worth building — agent-assisted product discovery on real signal. For us this is a serious encroachment on Notion/Confluence: Linear is already our project source of truth, and moving docs in next to issues could collapse two of our current sources into one. Worth a real look at the next tooling review (<a href="https://x.com/karrisaarinen/status/2062580986232373652">karrisaarinen</a>, <a href="https://x.com/karrisaarinen/status/2061506694124060673">launch</a>)</p><h3>SEO</h3><p><strong>Ahrefs’ billion-datapoint study reshapes AI-search strategy</strong> — Across 14 studies and 1B+ datapoints: “best X” listicles are 43.8% of all ChatGPT citations; 67% of top citations come from sources marketers can’t influence (Wikipedia 29.7%, homepages 23.8%); 28.3% of its most-cited pages have zero Google organic visibility — a separate discovery layer entirely. Schema markup had no meaningful impact. YouTube mentions had the highest correlation with AI brand visibility (0.737), beating backlinks and DR. AI Overviews now cut #1-result clicks by 58%. For WeRoad: YouTube-as-AI-SEO is the under-priced channel, and influencing AI citations needs a different playbook — listicles, video, being the source of truth — not schema tweaks (<a href="https://x.com/timsoulo/status/2061796432534003866">timsoulo</a>)</p><p><strong>Google Search Console is getting a dedicated AI performance report</strong> — The most-requested GSC report ever will cover both Search (AI Overviews, AI Mode) and Discover. The first release surfaces impressions only — not clicks — broken down by pages, countries, devices, and dates; queries are excluded for now, and rollout starts with a subset of sites. It’s the first first-party visibility into AI surface exposure, which until now has been a measurement black hole. Worth whitelisting WeRoad if Google offers opt-in, and building dashboards around impressions-without-clicks as a leading indicator of AI cannibalization (<a href="https://x.com/brodieseo/status/2062042288864244086">brodieseo</a>)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=47c2ae69b44e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Context Window — W22 2026]]></title>
            <link>https://medium.com/@smnbss/context-window-w22-2026-094e1f28973c?source=rss-5de037daf82f------2</link>
            <guid isPermaLink="false">https://medium.com/p/094e1f28973c</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai-sharing]]></category>
            <dc:creator><![CDATA[Simone Basso]]></dc:creator>
            <pubDate>Tue, 02 Jun 2026 07:11:49 GMT</pubDate>
            <atom:updated>2026-06-02T07:11:49.238Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>The “company brain” stopped being a concept and turned into a contested product category — in a single week.</strong> Nathan Baschez (founder of Lex) floated the spec out loud: agent-connectable shared context, nice UI, permissions, suggest-changes, versioning, SSO/RBAC, bundled connectors. 137K views, and a comments thread that reads like a market map — Glean, Notion AI, Outline+MCP, Cassidy, Mem, Brainiac (<a href="https://x.com/nbaschez/status/2059351438296584575">nbaschez</a>). Two days later Dhravya Shah open-sourced Supermemory — a universal company brain with native agent connectors, git-like versioning, RBAC, on-prem, six months of dogfooding behind it (<a href="https://x.com/DhravyaShah/status/2060429704449143159">DhravyaShah</a>). And Eric Siu landed the line that actually matters: the brain isn’t the folder, it’s the intelligence layer that sits between scattered context and the work the team needs done (<a href="https://x.com/ericosiu/status/2060435757127651664">ericosiu</a>, <a href="https://x.com/ericosiu/status/2060415100603781497">long-form teardown</a>). Even NotebookLM finally shipped live Drive sync so the brain stops drifting out of date (<a href="https://x.com/joshwoodward/status/2059665697618493706">joshwoodward</a>). Here’s the thing: this is exactly what we’re building in-house with Outline + Brainiac + the Vertex trial. The wiki is commoditised. The routing layer — what gets retrieved, by which agent, with what permissions — is the asset. Everyone now agrees on the problem. Almost nobody has shipped the boring enterprise plumbing that makes it real.</p><p><strong>The browser became the agent’s hands, and the tooling caught up all at once.</strong> Microsoft shipped Webwright, a Playwright-derived runtime built for agents where every session becomes a reusable, replayable workflow instead of being re-derived each run — and practitioners already invested in Playwright are eyeing it as a drop-in upgrade (<a href="https://x.com/mr_r0b0t/status/2059026191646945515">mr_r0b0t</a>, <a href="https://x.com/omarshahine/status/2059170907521003636">omarshahine</a>). Greg Brockman showed Codex fanning a single prompt into seven concurrent browser sessions — flights, car rental, lodging, forms, checkout, all in parallel (<a href="https://x.com/gdb/status/2059735815262249392">gdb</a>). Google promoted Chrome DevTools for agents to a stable 1.0, finally closing the write-verify loop: the agent can see console errors, network traffic, and layout instead of writing code blind (<a href="https://x.com/chromiumdev/status/2060114203621335523">ChromiumDev</a>). And Google Cloud’s Gemini Enterprise sandbox added computer use on top of code execution (<a href="https://x.com/rseroter/status/2060412856386879931">rseroter</a>). Three months ago “drive a browser with an LLM” was a flaky party trick. Now it’s a substrate with a replay engine, a parallelism story, and a built-in debugger. The browser is no longer something agents struggle to use — it’s the default surface they act through.</p><p><strong>Every hyperscaler now ships a managed agent runtime, and the moat moved to the control plane.</strong> Google Cloud’s Gemini Enterprise Agent Platform spins up per-task sandboxes with computer use and custom containers, isolating side effects from unsafe work (<a href="https://x.com/rseroter/status/2060412856386879931">rseroter</a>). A new GCP codelab walks deploying a BigQuery data-science agent — persistent memory across sessions — to Agent Engine for hosted scaling (<a href="https://x.com/GoogleCloudTech/status/2060360512672518262">GoogleCloudTech</a>). OpenAI killed the last enterprise blocker by letting ChatGPT, Codex, and the Responses API reach private MCP servers over outbound-only HTTPS — no public endpoint, no inbound firewall hole (<a href="https://x.com/openaidevs/status/2059703536825565499">OpenAIDevs</a>). And Mastra shipped an Agent Builder where devs author tools and non-devs assemble them into agents, with RBAC and ownership baked in (<a href="https://x.com/calcsam/status/2060377968334303467">calcsam</a>). The pattern is the same everywhere: sandboxing, hosted memory, RBAC, private connectivity. The runtime is becoming infrastructure. What you compete on isn’t whether you can run an agent — it’s governance, isolation, and who’s allowed to do what.</p><p><strong>The desktop assistant turned into a super-app war, and shipping velocity is the only weapon.</strong> Microsoft is reportedly redesigning Copilot into a “super app” built around an agent called Scout, repositioning the whole thing from chat window to agentic surface (<a href="https://x.com/alexeheath/status/2060516372891992097">alexeheath</a>). Google’s I/O recap bragged 100+ features in under 100 days since the April Gemini-for-macOS launch, plus a new “Spark” mode and a beefed-up voice stack (<a href="https://x.com/Google/status/2059752366195245075">Google</a>). And the way agents and humans collaborate is shifting too — bersabel’s demo of a shared live markdown canvas with charts and embeds is the same direction as Artifacts and Composer: the document becomes the conversation (<a href="https://x.com/brsabel/status/2060471698764521543">brsabel</a>). The consumer-AI race is now fully agent-first. Nobody’s winning on a smarter model. They’re winning on how fast they can wrap the model in a surface people actually live in.</p><p><strong>And the long tail confirms it: every horizontal SaaS gap is becoming an agent-callable vertical.</strong> Open Design 0.8.0 repositioned itself from canvas-first editor to a substrate where every asset, system, and prototype is a plugin an agent can invoke — the unit of reuse stops being the Figma component (<a href="https://x.com/nexudotio/status/2059549227538096378">nexudotio</a>). Notion added Gemini 3.5 Flash to its model picker, leaning into the multi-model posture we already run through LiteLLM (<a href="https://x.com/NotionHQ/status/2060459471445659866">NotionHQ</a>). Google’s open-source LangExtract is being pitched as matching $100K enterprise IDP tools for document extraction (<a href="https://x.com/mdancho84/status/2060319803454239117">mdancho84</a>). Linear wired Asks into Slack so intake happens by @-mention (<a href="https://x.com/linear/status/2059693937304932609">linear</a>). Different corners, same move: design, extraction, intake, KB — each one is being rebuilt around an agent endpoint. The “AI as a feature” era is over. The product surface itself is now the agent.</p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*dqV8G1ziMnthK4XtXkKvZQ.png" /></figure><h3>AI / Agents</h3><ul><li><strong>Nathan Baschez specs the “shared company brain for agents”</strong> — The Lex founder floated a product spec that’s effectively a company brain you connect to any agent: nice UI for viewing pages, permissions plus suggest-changes mode, versioning, SSO/RBAC, and bundled connectors to internal tools. 137K views and a comments thread that doubles as a market survey — Glean, Notion AI, Outline+MCP, Cassidy, Mem, Brainiac. The demand signal is loud and the field is fragmented, which means there’s a real startup-shaped gap for whoever ships the version with proper enterprise plumbing. Directly resonant with what we’re already building via Outline + gbrain (<a href="https://x.com/nbaschez/status/2059351438296584575">nbaschez</a>)</li><li><strong>Supermemory goes open source as a universal company brain</strong> — Dhravya Shah open-sourced Supermemory: native connectors for any agent, git-like versioning, RBAC, connectors to every common sync tool, and an on-prem deployment story, with six months of internal dogfooding behind it. It instantly becomes a credible OSS option to benchmark against the proprietary stack — relevant to WeRoad’s parallel Outline (KB) + Brainiac (Monkeys) + Vertex AI push. Worth a hands-on eval before we commit further down any single vendor path (<a href="https://x.com/DhravyaShah/status/2060429704449143159">DhravyaShah</a>)</li><li><strong>Eric Siu: the real brain is the intelligence layer, not the folder</strong> — Siu argues the company-brain conversation keeps missing the point — every company already has context scattered across calls, docs, Slack, dashboards, SOPs, and people’s heads. The moat isn’t another wiki; it’s the intelligence layer that sits between that context and the work the team needs done. A companion long-form teardown walks how Single Brain stitched Slack, Gong, and HubSpot together. This is the sharpest framing of the week and the one that maps cleanest onto our Outline-vs-Brainiac-vs-Vertex decision: tools are easy, the routing layer is the real bet (<a href="https://x.com/ericosiu/status/2060435757127651664">ericosiu</a>, <a href="https://x.com/ericosiu/status/2060415100603781497">long-form</a>)</li><li><strong>NotebookLM finally auto-syncs Google Drive</strong> — Josh Woodward (VP, Google Labs) announced NotebookLM’s top feature request: when source docs change in Drive, the notebook updates without a manual re-import. Rolling out at 10% of users with broader expansion soon. This closes one of the biggest gaps versus live-indexing tools like Glean, Cassidy, and Outline — and it’s a useful reference point as we evaluate Outline-as-company-KB and Vertex AI for the Monkeys stack (<a href="https://x.com/joshwoodward/status/2059665697618493706">joshwoodward</a>)</li><li><strong>Codex fans one prompt into seven parallel browser sessions</strong> — Greg Brockman highlighted a demo where Codex spawns concurrent browser-using sub-agents from a single user prompt — flight booking, car rental, lodging, hiking routes, forms, and checkout running in parallel, walked through in a 6-minute video. The striking part isn’t that one AI drives Chrome; it’s the fan-out into seven simultaneous sessions. This is OpenAI moving Codex from “code assistant” to general computer-use orchestrator, head-to-head with Anthropic’s claude-in-chrome and the Comet/Dia agentic browsers (<a href="https://x.com/gdb/status/2059735815262249392">gdb</a>)</li><li><strong>Microsoft ships Webwright — Playwright reimagined for agents</strong> — Microsoft released Webwright, a Playwright-derived runtime built specifically for agentic browser automation. The headline idea: every browser session becomes a reusable, replayable workflow, so agents capture and replay interactions instead of re-deriving them each run. Ships with a Nous Research Hermes Agent skill out of the box. If it delivers, it meaningfully lowers the cost and flakiness of letting LLM agents drive real browsers — and engineers already on Playwright are eyeing it as a drop-in upgrade rather than a new toolchain (<a href="https://x.com/mr_r0b0t/status/2059026191646945515">mr_r0b0t</a>, <a href="https://x.com/omarshahine/status/2059170907521003636">omarshahine</a>)</li><li><strong>OpenAI adds private MCP servers behind outbound-only HTTPS</strong> — ChatGPT, Codex, and the Responses API can now connect to private MCP servers without exposing them publicly: servers stay inside your network, the connection runs over outbound-only HTTPS, no inbound firewall hole, no public endpoint to harden. Ships with a dedicated “secure MCP tunnels” guide. For any enterprise that wanted MCP but couldn’t justify exposing internal tool servers, this removes the main blocker — and signals OpenAI pushing MCP as the canonical extension protocol (<a href="https://x.com/openaidevs/status/2059703536825565499">OpenAIDevs</a>)</li><li><strong>Gemini Enterprise sandbox adds computer use and custom containers</strong> — Google Cloud’s Gemini Enterprise Agent Platform now spins up per-task sandboxes supporting web browsing and custom containers on top of code execution. The sandbox is spawned by the agent for potentially unsafe tasks, isolating side effects. It lines up directly against AWS Bedrock AgentCore and Anthropic’s computer use — every hyperscaler now ships a managed agent runtime with sandboxed tool use as a first-class feature (<a href="https://x.com/rseroter/status/2060412856386879931">rseroter</a>)</li><li><strong>Mastra launches Agent Builder — internal agent platform with RBAC</strong> — Devs write tools, pick models, and build workflows; non-devs assemble those primitives into agents (or vibe-code them), chat with them, and publish. RBAC, ownership, and visibility controls included; everything compiles down to Mastra agents under the hood. This is the internal-agent-platform pattern every mid-size company will eventually need — useful prior art for what an in-house “WeRoad agents” platform looks like (<a href="https://x.com/calcsam/status/2060377968334303467">calcsam</a>)</li><li><strong>BigQuery data-science agent codelab → Agent Engine</strong> — A new Google Cloud codelab walks building a data-science agent that queries BigQuery public datasets, remembers user preferences across sessions with persistent memory, and ships to Agent Engine for hosted scaling. Directly relevant to the WeRoad analyst-agent stack on dbt_raw/serving — same pattern, hosted runtime. Worth a side-by-side against the in-house ADK setup before we decide what to run ourselves (<a href="https://x.com/GoogleCloudTech/status/2060360512672518262">GoogleCloudTech</a>)</li><li><strong>Microsoft plans Copilot super-app redesign with an agent called Scout</strong> — Alex Heath shared a leaked screenshot of Microsoft’s upcoming Copilot redesign, repositioned as a “super app” centered on an agent named Scout, with an announcement apparently imminent. The reframe is from chat window to agentic surface — Microsoft’s answer to Anthropic on the desktop and Google’s Gemini app push. The consumer-AI race is now fully agent-first, and the surface is the battleground (<a href="https://x.com/alexeheath/status/2060516372891992097">alexeheath</a>)</li><li><strong>Gemini for macOS: 100+ features in under 100 days</strong> — Google’s I/O recap: since the April Gemini-for-macOS launch, the team shipped 100+ features in under 100 days, adding “Gemini Spark” (not yet detailed) and a more powerful voice-understanding stack. It reads as Google’s answer to the ChatGPT and Claude desktop apps — the desktop assistant is now a contested surface, and shipping velocity is the lever everyone’s pulling (<a href="https://x.com/Google/status/2059752366195245075">Google</a>)</li><li><strong>Google LangExtract — open-source extraction “better than $100K tools”</strong> — Matt Dancho launched a thread on Google’s new open-source LangExtract, a document-extraction library he claims rivals enterprise OCR/IDP tools costing $100K+. Free, open source, Google-built. If it lives up to the hype it displaces Document AI, Unstructured.io, and a chunk of the IDP vendor stack — worth a quick eval for the contracts and expense-receipts pipeline (<a href="https://x.com/mdancho84/status/2060319803454239117">mdancho84</a>)</li><li><strong>Notion Custom Agents now run on Gemini 3.5 Flash</strong> — Notion added Gemini 3.5 Flash to its Custom Agents model picker, with hover-compare for speed, smarts, and cost across models. Notion is increasingly multi-model (Claude + GPT + Gemini) rather than locked to one provider — the same posture WeRoad runs through the LiteLLM proxy. “Let users pick the model” is becoming table-stakes UX for agent platforms (<a href="https://x.com/NotionHQ/status/2060459471445659866">NotionHQ</a>)</li><li><strong>Google adds collaborative Projects and Workflow Agents to Gemini for Business</strong> — Two additions land first in Gemini Enterprise before reaching the core business tier: collaborative Projects (shared workspaces, à la ChatGPT/Claude Projects) and Workflow Agents that automate cross-application tasks. Google catching up on the team-workspace pattern that’s become table stakes, with the workflow-agent piece as its answer to OpenAI’s Operator and Anthropic’s computer-use agents — at the productivity-suite layer rather than the raw API (<a href="https://x.com/testingcatalog/status/2059663483382227097">testingcatalog</a>)</li><li><strong>Coding agent with a shared live markdown canvas</strong> — bersabel demoed a coding-agent UX where the agent and human share a live markdown editor with interactive features — charts, color pickers, an HTML canvas, video embeds. The agent’s scratchpad becomes a rich collaboration surface rather than a chat log. Same direction as Cursor’s Composer and Claude’s Artifacts: the IDE becomes the document, and the document becomes the conversation (<a href="https://x.com/brsabel/status/2060471698764521543">brsabel</a>)</li><li><strong>Hermes Agent: self-evolving skills, three-tier memory, GEPA optimization</strong> — Akshay published a full masterclass on Hermes Agent, an open framework bundling SOUL.md personality files, a three-tier memory architecture, a Curator that consolidates skills, and a GEPA-based offline optimization pipeline — plus Telegram integration, multi-agent profiles, and cron automation. The framing emphasizes persistent self-improvement: agents that get measurably better across sessions rather than stateless function-callers. Useful reference architecture for the brain/super work (<a href="https://x.com/akshay_pachaar/status/2059644786857161106">akshay_pachaar</a>)</li></ul><h3>Engineering / DevOps</h3><ul><li><strong>Chrome DevTools for agents hits a stable 1.0</strong> — Google shipped Chrome DevTools for agents 1.0, exposing browser debugging, device emulation, and automated audits to AI coding assistants through a Chrome DevTools MCP server. It addresses one of the core blind spots of coding agents — they can write code but can’t observe runtime behavior, network traffic, console errors, or layout. With MCP-driven DevTools access, agents close the write-verify loop and self-correct on frontend work without a human in the middle. Strong fit for our gstack-qa pipeline, which already leans on the chrome-devtools MCP (<a href="https://x.com/chromiumdev/status/2060114203621335523">ChromiumDev</a>)</li><li><strong>Google Workspace Drops: Vids recording, Sheets auto-table, customizable meeting notes</strong> — The May “Workspace Drops” bundle ships AI features across the suite: Vids screen recording, Sheets “Convert to Table” auto-typing of imported CSVs, and customizable AI meeting notes in Meet. The meeting-notes customization is the standout — directly relevant to the brain’s meeting-harvest pipeline, since structured in-Meet digests could replace some of the Granola/manual steps (<a href="https://x.com/GoogleWorkspace/status/2060406566897992049">GoogleWorkspace</a>)</li></ul><h3>Design</h3><ul><li><strong>Open Design 0.8.0 turns assets, systems, and prototypes into agent-callable plugins</strong> — Open Design shipped v0.8.0, repositioning from a canvas-first editor to a substrate where every design artifact — assets, design systems, prototypes, exports, workflows — becomes a plugin an agent can invoke. The pitch is “from canvas-first to agent conversation”: instead of opening a file and dragging rectangles, you ask an agent and it composes from the plugins your team has published. This is the design-tooling version of the broader “everything becomes an MCP-shaped tool” pattern, and the most interesting bet for design ops — the unit of reuse stops being the Figma component and becomes the agent-callable plugin (<a href="https://x.com/nexudotio/status/2059549227538096378">nexudotio</a>)</li></ul><h3>Product Management</h3><ul><li><strong>Linear Asks now files requests from Slack via @-mention</strong> — Linear shipped a Slack integration for Asks: anyone can @-mention the bot, which picks the right intake template, asks for missing details conversationally, and creates the issue automatically — no more “can you file a ticket?” friction. It’s a familiar pattern (Jira Service Management, Halp), but Linear’s edge is the tight loop into its issue model. Relevant for any team running internal-requests-as-Linear-issues at WeRoad (<a href="https://x.com/linear/status/2059693937304932609">linear</a>)</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=094e1f28973c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Context Window — W21 2026]]></title>
            <link>https://medium.com/@smnbss/context-window-w21-2026-e2a344c758b5?source=rss-5de037daf82f------2</link>
            <guid isPermaLink="false">https://medium.com/p/e2a344c758b5</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai-sharing]]></category>
            <dc:creator><![CDATA[Simone Basso]]></dc:creator>
            <pubDate>Mon, 25 May 2026 09:59:56 GMT</pubDate>
            <atom:updated>2026-05-25T09:59:56.310Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>The “harness” stopped being a metaphor and became a product category.</strong> Three independent posts this week converged on the same word. Daniel San reported from running RAG over a large codebase that “the index was always one sprint behind the repo” and concluded you invest in the <em>harness</em> and let the agent traverse the live code, not in vectors (<a href="https://x.com/dani_avila7/status/2056187602181521584">dani_avila7</a>). Ronin posted the full pre-flight install order — direnv + a real secrets manager, LiteLLM/Portkey as the proxy, uv pinned per eval, mitmproxy on every call, inspect-ai for evaluations — and called it “the harness checklist” (<a href="https://x.com/DeRonin_/status/2056083767333163044">DeRonin_</a>). Then Chatbase crossed $10M ARR and re-launched itself with the literal pitch: “what Claude Code is to coding, Chatbase is to customer experience — the harness” (<a href="https://x.com/yasser_elsaid_/status/2056394356769366166">yasser_elsaid_</a>). Shopify’s Farhan Thawar made it a deadline: harness agents in 2026 or be behind (<a href="https://x.com/zodchiii/status/2056336049589092866">zodchiii</a>). The model has commoditised. The harness — proxy, evals, secrets, observability, guardrails, HITL — is now the moat. Pick yours and stop benchmarking models.</p><p><strong>Agent observability shipped, all in one Friday.</strong> On 2026–05–22, three platform vendors landed the same feature within hours of each other. Anthropic ships /usage in Claude Code with per-Skill, per-Agent, per-MCP, per-Plugin token breakdowns (<a href="https://x.com/bcherny/status/2057476878110261587">bcherny</a>). OpenAI ships Codex enterprise analytics — active users, credits, tokens, runs, leaderboards, lines of code, plugin usage, plus a matching Analytics API (<a href="https://x.com/openaidevs/status/2057530213974814844">openaidevs</a>). Notion ships Custom Agent Insights — full visibility into every run of every custom agent (<a href="https://x.com/notionhq/status/2057502296196714759">notionhq</a>). Google Chrome bolted on the same idea at the runtime layer, demoing DevTools-for-agents that auto-verifies and auto-fixes flows using console + network signals (<a href="https://x.com/chromiumdev/status/2057172959333171500">chromiumdev</a>). Three months ago &quot;agent observability&quot; was a vague buzzword. Now it&#39;s the spec sheet every platform admin reads first. The agent war in 2026 is being fought on governance and transparency, not raw capability.</p><p><strong>Local computer-use crossed from demo to shippable.</strong> Google’s Gemma team posted a clip of Gemma 4 E4B — an on-device open model — driving an iOS simulator end-to-end via Argent. Complex UI navigation, fully local, no cloud loop (<a href="https://x.com/googlegemma/status/2057570113390551452">googlegemma</a>). The same week, leaks showed the Gemini desktop app splitting into two workspaces: standard Chat and a new <strong>Spark</strong> mode for local agentic tasks — folder access, code analysis, script generation, Drive sync, plus a “Stream to Cursor” feature that reads context from whichever app your cursor hovers over (<a href="https://x.com/pankajkumar_dev/status/2056380438801953223">pankajkumar_dev</a>). Google is now racing Anthropic Computer Use and OpenAI Operator for the desktop-agent surface — and it’s the first time the open-weights side has a credible local entry. The next six months will be loud. Privacy-preserving mobile and desktop agents go from research demo to product category.</p><p><strong>Generative AI is breaking out of the chat box.</strong> Anthropic teased “Claude Design”, a prototype turning Claude into a design tool. Details thin, engagement massive (754K views) — the appetite is obvious (<a href="https://x.com/claudeai/status/2057487463518548073">claudeai</a>). ElevenLabs shipped Speech Engine: drop a voice layer onto any existing chatbot with one prompt, 70+ languages, 11,000+ voices, no LLM swap required (<a href="https://x.com/dr_cintas/status/2057490783045841354">dr_cintas</a>). Okara pitched itself as the “first SEO Agent” — full-funnel autonomous SEO that audits, fixes, and publishes optimised content (<a href="https://x.com/askokara/status/2057476509099540556">askokara</a>). Trevin Chow launched Printing Press Library, an npm registry of agent-callable CLIs across 17 categories — the skill-aware package manager that pip and npm haven’t shipped yet (<a href="https://x.com/trevin/status/2057664119357227211">trevin</a>). The pattern is the same in every direction: design surfaces, voice surfaces, SEO surfaces, package registries — every horizontal SaaS gap is now an agentic vertical. The “AI as feature” era ends here. The “AI as the product surface” era is the one we’re building in.</p><p><strong>Workspace platforms are converging on the same playbook.</strong> Gemini Enterprise rolled out Shared Chats, Projects, and a Canvas experience for Docs inside Gemini — multi-user conversations against shared context, persistent workspace containers, the artifact as the durable unit of work (<a href="https://x.com/GoogleWorkspace/status/2056480700220448954">GoogleWorkspace</a>). Google AI Studio added native Workspace tool access — pull from Sheets, organise Drive, process docs without leaving Studio (<a href="https://x.com/googleaistudio/status/2056845028136489419">googleaistudio</a>). Combined with Notion’s Agent Orchestration from last week, the company-KB layer has a shape: shared context, multi-agent participation, humans-in-the-loop for approvals, and the document — not the chat — as the persistent state. This is exactly the territory Cass is rolling out for WeRoad through Outline + Vertex AI. The race for the AI-native workspace is on, and the differentiator is no longer the model — it’s how cleanly your knowledge graph hands off to the next agent.</p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*dqV8G1ziMnthK4XtXkKvZQ.png" /></figure><h3>AI / Agents</h3><p><strong>Claude Code </strong><strong>/usage — finally, per-skill token observability</strong> — The next Claude Code build ships a /usage slash command that breaks down token consumption by Skill, Agent, MCP, and Plugin. CLI first, Claude Desktop next. Power users running stacks of skills and agents in parallel have been guessing which plugin is eating their context bill — this ends the guessing game. The bigger signal is that &quot;what is burning my budget&quot; is now the #1 friction point at the top of the user funnel, not a power-user edge case (<a href="https://x.com/bcherny/status/2057476878110261587">bcherny</a>)</p><p><strong>OpenAI ships Codex enterprise analytics — admin-grade observability</strong> — OpenAI upgraded Codex analytics for business/enterprise plans: breakdowns of active users, credits, tokens, runs, user leaderboards, lines of code generated, plugin usage, plus matching Analytics API updates. Mirrors Anthropic’s /usage on the same day. For CTOs rolling out coding agents at scale, this is exactly the governance plumbing that was missing — you can finally tell who is using what, how much, and on which projects. Same convergence story as Notion&#39;s Insights and Anthropic&#39;s /usage: the platform war moved from capability to control plane (<a href="https://x.com/openaidevs/status/2057530213974814844">openaidevs</a>)</p><p><strong>Notion Custom Agent Insights — every run, every action, every error</strong> — Notion shipped Custom Agent Insights with full visibility into every run of every custom agent: what it did, what it touched, where it went wrong. Third “agent observability” launch in a single day (alongside Anthropic and OpenAI). For non-engineering teams adopting agents inside Notion, this is the safety net that was missing — you can now run customer-facing or finance-touching agents without a black box. Confirms the category-wide thesis: in 2026, agent platforms compete on transparency (<a href="https://x.com/notionhq/status/2057502296196714759">notionhq</a>)</p><p><strong>Anthropic teases Claude Design — first-party generative design tool</strong> — Anthropic posted a short teaser of “Claude Design”, a prototype experiment turning Claude into a design tool. Details thin, engagement massive — 754K views, 7.6K likes, 2.3K bookmarks on a teaser. Signals Anthropic is following the same playbook OpenAI and Figma have run into the design-tool layer, likely powered by the new Artifact and Computer Use primitives. Likely public R&amp;D rather than imminent launch, but the audience response says the demand is real and unmet (<a href="https://x.com/claudeai/status/2057487463518548073">claudeai</a>)</p><p><strong>Gemma 4 E4B drives an iOS simulator autonomously via Argent</strong> — Google’s Gemma team showed Gemma 4 E4B — an open, on-device model — navigating an iOS simulator end-to-end with Argent as the agentic runtime. Complex UI interactions and software navigation handled locally, no cloud loop. The framing is “a new era of on-device automation”. Local computer-use is no longer science fiction; expect a wave of privacy-preserving mobile agents in the next 6–12 months. Significant because it’s the first credible open-weights entry in a category dominated by Anthropic Computer Use and OpenAI Operator (<a href="https://x.com/googlegemma/status/2057570113390551452">googlegemma</a>)</p><p><strong>Gemini desktop leaks — Chat + Spark modes, Stream to Cursor, local Skills</strong> — Pre-Google-I/O leaks indicate the Gemini desktop app is splitting into two workspaces: standard <strong>Chat</strong> mode and <strong>Spark</strong> mode for local agentic tasks (folder access, code analysis, scripts, file organisation, Drive sync). A new “Stream to Cursor” feature reads context from whichever app/window the cursor hovers over; a floating overlay can share screen/window/camera context. Internally “Gemini Omni” is referred to as “Veo4 Omni”, and local “Skills” support is in the works. Google racing the same desktop-agent surface as Anthropic and OpenAI — and bringing the deepest OS-integration story of the three (<a href="https://x.com/pankajkumar_dev/status/2056380438801953223">pankajkumar_dev</a>)</p><p><strong>Gemini Enterprise — Shared Chats, Projects, Canvas-as-Docs</strong> — Demo walkthrough from Google Cloud Next of three new Gemini Enterprise primitives: Shared Chats (multi-user conversations against the same context), Projects (persistent workspace containers), and a new Canvas experience for Docs <em>inside</em> Gemini Enterprise. Framing: “stop working in silos, start executing in a single pane of glass.” Directly comparable to Notion AI and Claude Projects but with multi-agent participation as a first-class concept. Worth a hard look as we evaluate Vertex AI for the WeRoad workspace — pairs with the Outline rollout Cass is running (<a href="https://x.com/GoogleWorkspace/status/2056480700220448954">GoogleWorkspace</a>)</p><p><strong>Google AI Studio adds native Workspace tool access</strong> — You can now build apps inside Google AI Studio that directly call Workspace: pull from Sheets, organise Drive, process team docs — without leaving the Studio environment. Natural follow-on to the Vertex AI trial: prototyping Workspace-grounded internal tools (HR forms, finance scrapers, OKR roll-ups) becomes a near-no-code exercise inside AI Studio. The “build agent, ground in Workspace, ship” loop is finally one product surface. Worth pinging the People and Finance teams (<a href="https://x.com/googleaistudio/status/2056845028136489419">googleaistudio</a>)</p><p><strong>ElevenLabs Speech Engine — drop voice onto any chatbot, no LLM swap</strong> — ElevenLabs shipped Speech Engine: a voice layer that drops on top of an existing chat agent without touching the LLM, RAG, or architecture. One prompt, 70+ languages, 11,000+ voices, combining speech, transcription, and voice orchestration into a single pipeline. For voice-AI use cases (think the WeRoad inbound call drop diagnostic we ran), this is the cleanest path yet from text-only chatbot to a phone-ready voice agent. The voice layer is finally decoupled from the model layer (<a href="https://x.com/dr_cintas/status/2057490783045841354">dr_cintas</a>)</p><p><strong>Printing Press Library — npm registry for agent-native CLIs</strong> — Trevin Chow (ex-Disney+/Hulu eng leadership) launched the Printing Press Library: a unified registry of hundreds of agent-native CLIs across 17 categories, contributed by the @ppressdev community. npm i -g @mvanhorn/printing-press-library gets you a search + install command that pulls down both the binary and its agent skill in one shot. Early days (1.2K views), but the pattern — a skill-aware package manager for the agent era — is exactly what npm/pip don&#39;t yet do natively. Watch this slot land (<a href="https://x.com/trevin/status/2057664119357227211">trevin</a>)</p><p><strong>Codex </strong><strong>/goal — persistent-objective mode goes mainstream</strong> — Greg Brockman demoed a new /goal slash command in Codex that pins a long-running objective and keeps the agent iterating until it&#39;s actually solved, rather than stopping at the first plausible-looking output. Mirrors Anthropic&#39;s /goal from last week and the broader &quot;ralph-loop / persistent goal&quot; pattern emerging across coding harnesses (Claude Code, Cursor, Aider). Persistent-goal mode is becoming a default expectation, not a power-user trick. Don&#39;t ship a coding agent without it (<a href="https://x.com/gdb/status/2056430780809892252">gdb</a>)</p><p><strong>Chatbase hits $10M ARR — “the harness for customer-facing agents”</strong> — Chatbase crossed $10M ARR and re-launched as a full harness for customer-facing AI agents, explicitly framed as “what Claude Code is to coding, Chatbase is to customer experience”. The stack: model + context + tools + workflows + guardrails + human-in-the-loop, going beyond ticket deflection toward brand ambassadorship across every channel. The “harness” framing is becoming the standard vocabulary for agent platforms. Worth tracking as the customer-care reference architecture as we think about WeRoad’s own Zendesk-side agent stack (<a href="https://x.com/yasser_elsaid_/status/2056394356769366166">yasser_elsaid_</a>)</p><p><strong>Shopify’s eng head: “harness agents in 2026 or be behind”</strong> — Interview clip with Shopify’s Head of Engineering Farhan Thawar laying out the full Shopify AI playbook — “the most practical breakdown of enterprise AI coding I’ve seen this year.” Headline framing: if you don’t figure out how to harness agents in 2026, you’ll be behind. Shopify is one of the most quoted enterprise AI-engineering programs and a useful peer benchmark — especially for the Plenaria Tech narrative. The vocabulary again: harness, not model (<a href="https://x.com/zodchiii/status/2056336049589092866">zodchiii</a>)</p><p><strong>The pre-flight stack for any new agentic project</strong> — Ronin posted the full opinionated install order before starting any agentic project. (1) direnv + a real secrets manager (1Password CLI, Doppler, Infisical, Vault) — scope keys per folder, kill the “API key in .env” breach pattern. (2) LiteLLM or Portkey as a single proxy in front of every model — caching (30–60% bill cut), automatic 429 fallback, per-feature/per-user budget caps, PII redaction. (3) uv + git commit on every passing eval, pinning prompt + code + model version + pass rate together. (4) mitmproxy in front of every LLM call to see prompts, retries, and prompt-injection moments live. (5) inspect-ai (the eval framework Anthropic, DeepMind, UKAISI actually use) plus a /lessons.md per repo. Lift this directly into our WeRoad platform docs (<a href="https://x.com/DeRonin_/status/2056083767333163044">DeRonin_</a>)</p><p><strong>RAG-on-code fails — invest in the harness, not the index</strong> — Short field report from running RAG with semantic search and vector DBs over large codebases at a previous startup: “the index was always one sprint behind the repo.” Same shift Cursor and Claude Code popularised — invest in the harness and let the agent traverse the live codebase rather than maintaining a parallel embedding index. Reinforces the “give the model real tools, not stale vectors” pattern that’s now consensus for code-aware agents. Don’t bother with a code vector DB. Build the harness instead (<a href="https://x.com/dani_avila7/status/2056187602181521584">dani_avila7</a>)</p><p><strong>Hermes Agent — full operator stack on one VPS</strong> — Long-form X article on operating Hermes Agent — the open-source agent framework by Nous Research, currently #1 on OpenRouter for global token usage. Walks the agent control-room template, configuring specialist agents (brain / personality / skillset all in one folder), and scaling from one agent to a “marketing company on one VPS.” Hermes browses, runs terminal commands, schedules cron jobs, monitors inboxes, drafts work, and posts results to Telegram/Discord/Slack/email. The clearest reference for how solo operators are stitching together always-on agent stacks today (<a href="https://x.com/shannholmberg/status/2055335043904492011">shannholmberg</a>)</p><p><strong>Google CodeWiki — paste a GitHub URL, get interactive docs + Gemini chat</strong> — Google released CodeWiki, open source, which takes any GitHub repo and turns it into interactive documentation with architecture diagrams and a Gemini chat over the code. DeepWiki-style experience, but native Google, free, and self-hostable. Worth a head-to-head comparison against DeepWiki for the outputs/services/*.AGENT.MD pipeline — if CodeWiki produces good architecture diagrams automatically, it could replace or augment our manual service docs (<a href="https://x.com/HowToAI_/status/2056435847311061287">HowToAI_</a>)</p><p><strong>“free-claude-code” — unlimited Claude Code via NVIDIA NIM proxy</strong> — Open-source project that lets you point Claude Code at localhost, where a proxy converts Anthropic API calls into NVIDIA NIM format using a free NVIDIA API key — 40 req/min, no subscription. Supports Kimi K2, GLM 4.7, MiniMax M2, Devstral, and others; streams thinking tokens and tool calls live; includes a Telegram bot for controlling Claude Code from your phone. Two-minute setup, 100% OSS. Useful escape hatch for personal experiments — not production, where you&#39;d want SLAs (<a href="https://x.com/dr_cintas/status/2056077614113325373">dr_cintas</a>)</p><p><strong>Cheat sheet — agentic AI architecture in 7 layers</strong> — A concise reference deck for designing or explaining agentic systems: define the goal, then stack Orchestration (flow &amp; coordination), Agents (specialised workforce), Tools (APIs, search, DBs), Memory (short + long-term), Monitoring, Reliability &amp; Failure (retries, fallbacks, HITL), and Governance &amp; Security (auth, compliance, audit). Punchline: “agents alone don’t make systems powerful — architecture does.” Useful as a mental checklist when reviewing an internal agent design for missing layers (<a href="https://x.com/AamirAnsar94694/status/2056280345797087740">AamirAnsar94694</a>)</p><h3>Engineering / DevOps</h3><p><strong>Chrome DevTools for agents — closed-loop debug, no human required</strong> — Google captured a walkthrough at #GoogleIO showing Chrome DevTools for agents using console logs and network traffic to automatically verify — and then fix — a broken signup flow in real time. Pitch: “give your AI the runtime visibility it needs.” Pairs with the existing chrome-devtools MCP we already use for QA; the new angle is closed-loop agent debugging — observe → hypothesise → patch → re-verify — without a human in the loop. Strong fit for our gstack-qa pipeline. The runtime is no longer opaque to the agent (<a href="https://x.com/chromiumdev/status/2057172959333171500">chromiumdev</a>)</p><h3>SEO</h3><p><strong>Okara — “world’s first SEO Agent”, full-funnel autonomous SEO</strong> — Okara is pitching itself as the first end-to-end SEO Agent: feed it a URL, it audits 100+ data points, fixes issues, finds keyword gaps, and publishes optimised content to grow organic traffic — all autonomously. Whether it earns the “first” claim (it doesn’t — many startups have similar pitches), the agentic-SEO category is consolidating fast. Direct competitive read for WeRoad’s SEO/GEO motion. Worth a hands-on test against the existing seo-geo workflow before someone else gets there first (<a href="https://x.com/askokara/status/2057476509099540556">askokara</a>)</p><h3>Others</h3><p><strong>Tolaria adds optional “no Git” sync mode</strong> — New release of Tolaria (markdown editor by Luca Rossi of <em>Refactoring</em>) lets users disable and hide Git completely — for those who sync via iCloud, Dropbox, or similar. Ships with the usual round of bug fixes and performance tweaks. A “big thing, not mobile” is teased for Wednesday — keep an eye on it (<a href="https://x.com/lucaronin/status/2056308842506375657">lucaronin</a>)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e2a344c758b5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Context Window — W20 2026]]></title>
            <link>https://medium.com/@smnbss/context-window-w20-2026-083a751529a3?source=rss-5de037daf82f------2</link>
            <guid isPermaLink="false">https://medium.com/p/083a751529a3</guid>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Simone Basso]]></dc:creator>
            <pubDate>Wed, 20 May 2026 13:35:53 GMT</pubDate>
            <atom:updated>2026-05-20T13:35:53.586Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Hermes is the open-source agent OS now.</strong> Nous Research strung together a relentless cadence this week. Monday a native desktop app with persistent memory and multi-agent management (<a href="https://x.com/intheworldofai/status/2053362763842371586">intheworldofai</a>). Thursday background Computer Use on macOS, Kanban orchestration, Lightpanda backend, free Qwen3.6-Plus, long-term memory (<a href="https://x.com/intheworldofai/status/2054430265368273330">intheworldofai</a>). Friday Akshay’s masterclass dissecting the full stack: a SOUL.md identity layer, a three-tier memory system, 687 skills across six execution backends, and GEPA’s offline skill evolution at $2–10 a run (<a href="https://x.com/akshay_pachaar/status/2054564519280804028">akshay_pachaar</a>). 2M+ views on the deep dive. Stop waiting for “the open Claude.” It shipped. The SOUL/memory/skills triad is the most replicable blueprint in the wild — if you’re picking an agent OS pattern to study, study this one.</p><p><strong>Tool definitions belong in code, not in context.</strong> Akshay’s Monday breakdown made the math unignorable (<a href="https://x.com/akshay_pachaar/status/2053380764805583212">akshay_pachaar</a>): Playwright MCP 13.7K tokens to load, Chrome DevTools 18K, a five-server stack 55K before any work, a Drive → Salesforce flow ballooning to 150K because the transcript got piped through the model twice. Code Mode flips it — the agent writes TypeScript that imports only the tools it needs. Same task drops to 2K tokens. 98.7% reduction. Cloudflare collapsed 2,500 endpoints into 1K tokens by exposing just search and execute. Anthropic&#39;s MCP SDK downloads went from 100M to 300M this year. MCP isn&#39;t dying — the load-everything-upfront pattern is. Printing Press is already shipping the new shape: agent-callable CLIs distributed as npm packages (<a href="https://x.com/pejmanjohn/status/2053500526084727295">pejmanjohn</a>). Plan your stack around code, not schemas.</p><p><strong>Every platform is racing for the agent control plane.</strong> Anthropic shipped Claude Code’s unified agent view so fan-out sessions stop disappearing (<a href="https://x.com/claudeai/status/2053940934736228454">claudeai</a>) and a /goal slash command to keep agents on task (<a href="https://x.com/claudedevs/status/2054351031279186040">claudedevs</a>). Google launched the Agent Registry as the governance primitive for internal agent sprawl (<a href="https://x.com/googlecloudtech/status/2053550642761744569">googlecloudtech</a>) and pushed 50+ managed MCP endpoints to GA or preview (<a href="https://x.com/googlecloudtech/status/2053958329752322514">googlecloudtech</a>). Notion shipped Agent Orchestration so a Decagon ticket can hand off to a coding agent that pings a human before merging (<a href="https://x.com/notionhq/status/2054601077346246695">notionhq</a>), plus a developer platform with humans and coding agents as co-equal consumers (<a href="https://x.com/notiondevs/status/2054591579076403467">notiondevs</a>). YC W26 spun out Emdash (<a href="https://x.com/arnestrickmann/status/2054315390252327220">arnestrickmann</a>). Superset paired a CLI with its GUI (<a href="https://x.com/superset_sh/status/2054300997393248712">superset_sh</a>). The category has converged on one thesis: humans need a cockpit to drive many agents. Pick where yours lives before the agent count gets away from you.</p><p><strong>Programmatic agent usage is finally a line item.</strong> Anthropic announced that paid Claude plans will include a dedicated monthly credit for programmatic usage starting June 15 — Agent SDK, claude -p headless mode, GitHub Actions, and third-party Agent SDK apps all covered separately from interactive Claude Code (<a href="https://x.com/claudedevs/status/2054610152817619388">claudedevs</a>). Until now any CI run, scheduled job, or background automation ate the same quota as your live coding sessions, which made it impossible to budget. Treating &quot;agents as a line item&quot; sounds dull — it isn&#39;t. It&#39;s the signal that this stops being a hobbyist phase. You can plan agent workloads now. That changes what eng teams are willing to wire up.</p><p><strong>Incumbents are torching brand equity to escape the old map.</strong> Eoghan McCabe announced Intercom is renaming the parent company to Fin, after the AI customer-service agent they shipped three years ago (<a href="https://x.com/eoghan/status/2054238946612367818">eoghan</a>). All 1,400 employees now work for Fin. Intercom (the help-desk) keeps its name and just shipped Intercom 2, but Fin is about to be the largest part of the business. McCabe’s thesis: humans rarely update their mental map of incumbents, so newer entrants in AI categories beat “provably inferior technology” simply because they carry no baggage. Paperclip made the same move at smaller scale, dumping the maximalist “zero human companies” pitch for “the app people use to manage agents at work” (<a href="https://x.com/dotta/status/2054386037380751475">dotta</a>). The playbook is becoming visible: in an AI-native era, your existing brand is a liability. Burn it before a smaller competitor does.</p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*dqV8G1ziMnthK4XtXkKvZQ.png" /></figure><h3>AI / Agents</h3><p><strong>Code Mode ends the MCP-vs-CLI debate</strong> — Akshay’s deep dive reframes 2025’s MCP-vs-CLI argument: the protocol was never the problem, the habit of dumping every tool’s schema into model context was. Concrete receipts — Playwright MCP 13.7K tokens, Chrome DevTools 18K, a five-server stack 55K before any work, a Drive → Salesforce workflow ballooning to 150K because the transcript got piped through twice. Anthropic’s November paper flips the model’s job: the agent writes TypeScript that imports only the tools it needs. Same task drops to 2K tokens (98.7% reduction). Cloudflare collapsed 2,500 endpoints into 1K tokens with search and execute. MCP SDK downloads tripled to 300M this year — MCP isn&#39;t dying, the upfront tool-load pattern is. Tool definitions belong in code (<a href="https://x.com/akshay_pachaar/status/2053380764805583212">akshay_pachaar</a>)</p><p><strong>Hermes Masterclass — a full architectural tour of Nous Research’s agent OS</strong> — Akshay’s long-form X article walks the entire Hermes stack: SOUL.md as identity layer, three-tier memory (Markdown + full-text session search + external providers), self-evolving skills curated by a 30-day garbage collector, GEPA evolutionary skill refinement via execution traces at $2–10 per run, Telegram integration, profile creation, cron scheduling for autonomous daily jobs. 687 skills across 18 categories (87 built-in, 79 optional, 16 from Anthropic, 505 from LobeHub), six execution backends (local, Docker, SSH, Modal, Daytona, Singularity), a 90-turn hard cap to prevent runaway costs. The most studied open-source alternative to closed agent stacks right now (<a href="https://x.com/akshay_pachaar/status/2054564519280804028">akshay_pachaar</a>)</p><p><strong>Hermes ships multi-agent + Computer Use + persistent memory in one week</strong> — Nous Research stacked a fast burst of upgrades: WebUI/desktop app, background Computer Use on macOS, multi-agent orchestration, Kanban upgrades, Lightpanda browser backend, free Qwen3.6-Plus in Nous Portal, better autonomous workflows, persistent long-term memory. The framing has shifted — Hermes now feels less like an AI tool and more like an open-source agentic operating system. For teams wary of closed-source agent stacks (Claude Agent SDK, OpenAI Operator), Hermes is rapidly becoming the credible OSS alternative — worth tracking as a benchmark for what an open agent OS actually looks like (<a href="https://x.com/intheworldofai/status/2054430265368273330">intheworldofai</a>)</p><p><strong>Hermes Agent Desktop ships as a native app</strong> — Nous Research’s self-evolving 24/7 autonomous agent left the CLI behind and shipped as a full native desktop app. The bundle: multi-agent management, persistent memory, autonomous workflows, local infrastructure support, all in one UI. The “autonomous local agent” category is moving from terminal-first toolkits to packaged desktop products — Hermes joining a small cohort (Claude Desktop, OpenClaw) targeting operator-style use cases without cloud orchestration. Signal that local-first, always-on agents are crossing from research projects into shippable consumer surfaces (<a href="https://x.com/intheworldofai/status/2053362763842371586">intheworldofai</a>)</p><p><strong>Notion ships Agent Orchestration across the workspace</strong> — Notion announced Agent Orchestration: agents inside Notion can call other agents and loop humans in for approval mid-flow. Their example — a Decagon support ticket routes into Notion, which hands it to a coding agent that proposes a fix and pings the team to sign off before merging. Orchestration is no longer engineer-only; non-technical teams can wire agent flows directly inside docs and databases. Positions Notion as the human-in-the-loop layer between specialized agents (support, coding, ops) rather than just a document store. The artifact, not the chat, becomes the durable unit of work (<a href="https://x.com/notionhq/status/2054601077346246695">notionhq</a>)</p><p><strong>Notion launches a developer platform for agent builders</strong> — Alongside Agent Orchestration, Notion released a developer platform with primitives for syncing arbitrary data sources, building custom tools, and orchestrating multi-agent workflows — explicitly designed for both humans and coding agents to consume. The framing (“you and your coding agents”) signals Notion is treating agent-authored integrations as a first-class use case, not just human-built ones. Same pattern Slack and Linear are converging on: stop being an app, become an agent surface. For knowledge-work platforms, this is the new posture (<a href="https://x.com/notiondevs/status/2054591579076403467">notiondevs</a>)</p><p><strong>Claude Code adds Agent View — one list of all sessions</strong> — Anthropic shipped a research-preview “agent view” in Claude Code: a single unified list of all sessions. Claude Code already supported background tasks, subagents, and worktrees, but discovering and switching between them was scattered. One canonical session list gives users a control plane for fan-out workflows. Meaningful step toward Claude Code behaving like a long-running multi-agent orchestrator rather than a single-shot REPL. The discoverability gap was the real bottleneck for sustained agentic work — this closes it (<a href="https://x.com/claudeai/status/2053940934736228454">claudeai</a>)</p><p><strong>Claude Code ships </strong><strong>/goal to keep the agent on task</strong> — Anthropic shipped /goal, a slash command for declaring an explicit objective the agent should pursue until completion. Framed as one of several mechanisms — alongside autonomy controls and session features — to prevent Claude from drifting off-task or stopping prematurely on long jobs. The launch tweet pulled outsized attention in hours, a clear signal that &quot;keep the agent driving&quot; is the dominant pain point users want solved. Combined with the new agent view, Anthropic is layering control-plane primitives onto Claude Code for sustained agentic work (<a href="https://x.com/claudedevs/status/2054351031279186040">claudedevs</a>)</p><p><strong>Anthropic adds programmatic-usage credit to paid Claude plans</strong> — Starting June 15, paid Claude plans will include a dedicated monthly credit specifically for programmatic usage. The credit covers the Claude Agent SDK, the claude -p headless mode, Claude Code GitHub Actions, and any third-party app built on top of the Agent SDK. Until now, programmatic usage shared the same quota as interactive Claude Code sessions, which made budgeting agent workloads impossible. The change makes &quot;agents as a line item&quot; first-class — you can now plan automations and CI workflows without burning your interactive coding allowance (<a href="https://x.com/claudedevs/status/2054610152817619388">claudedevs</a>)</p><p><strong>Google Agent Registry inside Gemini Enterprise</strong> — Google Cloud announced an Agent Registry as part of its Gemini Enterprise Agent Platform — a single source of truth indexing every internal agent, tool, and skill across an organization. The pitch: simplify discovery and ensure only governed, approved assets surface to enterprise users. Google’s answer to the enterprise-agent sprawl problem that Salesforce, Microsoft, and ServiceNow are racing to solve. A registry layer is the natural governance primitive once an org has more than a handful of agents in production. Reference design for when internal agent catalogs become unavoidable (<a href="https://x.com/googlecloudtech/status/2053550642761744569">googlecloudtech</a>)</p><p><strong>50+ Google-managed MCP servers now GA or preview</strong> — Google Cloud now offers more than 50 managed MCP server endpoints, GA or in preview, with more on the way. Pointing an AI agent at a Google-managed MCP URL means inheriting the Google Cloud security stack — IAM, audit logging, network controls — without per-region configuration. Google’s answer to “every team rolls its own MCP server” sprawl: centralize the integration surface, ride on existing GCP trust boundaries. For enterprise agentic deployments, managed MCP dramatically lowers the security review burden (<a href="https://x.com/googlecloudtech/status/2053958329752322514">googlecloudtech</a>)</p><p><strong>Gemini Enterprise adds project workspaces for team + agent collaboration</strong> — Gemini Enterprise now supports “projects” — persistent shared spaces where humans and agents collaborate on a single body of work with bound data, shared conversations, and a collaborative canvas. Positions Gemini Enterprise closer to Notion-AI, ChatGPT Projects, and Claude Projects, but with multi-agent participation as a first-class concept. The collaborative canvas suggests Google is treating the artifact (not the chat) as the durable unit of work. Important for orgs trying to make agents stick beyond one-off prompts (<a href="https://x.com/googlecloud/status/2053913029788553459">googlecloud</a>)</p><p><strong>Google publishes 13 official Skills for Claude Code, Antigravity, Gemini CLI</strong> — Google opened a google/skills GitHub repo with 13 first-party Skills covering integrations with Claude Code, Antigravity, and Gemini CLI. Skills — Anthropic&#39;s format originally — are reusable agent capability packs, and Google releasing official ones signals the format is becoming a cross-vendor standard for distributing agentic recipes. The 13 published packs likely target Google Cloud workflows (GCP, BigQuery, Workspace) so engineers can drop in productized expertise without writing their own glue. Watch this as the MCP-server vs. Skill-pack debate heats up (<a href="https://x.com/commte/status/2054136870016356408">commte</a>)</p><p><strong>Crabbox pattern — Codex reproduces bugs in 10 parallel ephemeral sandboxes</strong> — Peter Steinberger describes his current bug-investigation loop: hand the bug to Codex, which spins up an ephemeral “crabbox” (crabbox.sh), recreates the exact failing state inside it, verifies the bug reproduces, fixes it, verifies the fix — all without touching his local machine. Two wins: no contamination from a polluted local env, and ~10 sessions run in parallel because each one is isolated. Concrete data point for the “agent-driven, sandbox-per-task” workflow that’s becoming the default for AI-native debugging — pattern eng teams should evaluate as Codex/Claude Code adoption grows (<a href="https://x.com/steipete/status/2053032450138276274">steipete</a>)</p><p><strong>Emdash — open-source Agentic Development Environment (YC W26)</strong> — Arne Strickmann is building Emdash, pitched as an “open-source Agentic Development Environment” and going through YC W26. The product positions itself as a workspace for orchestrating coding agents — analogous to what Superset and Paperclip are also chasing. With multiple YC-backed teams now racing to own the “agent IDE” category in early 2026, the bet is that humans will need a dedicated cockpit for managing many parallel coding agents rather than embedding them inside existing IDEs. Open source is the differentiator here (<a href="https://x.com/arnestrickmann/status/2054315390252327220">arnestrickmann</a>)</p><p><strong>Superset ships CLI — drive the agent orchestrator from the terminal</strong> — Superset (an editor for orchestrating coding agents) released a CLI that lets you create workspaces, spin up agents, and edit tasks from the terminal with the GUI updating in real time. Works with automation and remote workspaces, and includes a one-liner to teach the CLI to your agents as skills: npx skills add https://superset.sh. Installation: brew install superset-sh/tap/superset. The pattern — GUI for state, CLI for automation, skills for agent self-onboarding — is becoming the standard layout for the new agent orchestrator category (<a href="https://x.com/superset_sh/status/2054300997393248712">superset_sh</a>)</p><p><strong>Paperclip repositions: from “zero human companies” to “manage agents at work”</strong> — Paperclip’s founder dotta announced a positioning shift away from the maximalist “zero human companies” framing toward “the app people use to manage agents at work.” Deliberate move from a provocative end-state narrative to a concrete, sellable wedge: agent management as a daily-driver tool, not an existential bet. The pivot mirrors the broader category convergence (Superset, Emdash, Claude Code agent view) on “humans need a UI to drive many agents” rather than “agents replace humans.” Useful framing reference for any internal agent tooling positioning conversation (<a href="https://x.com/dotta/status/2054386037380751475">dotta</a>)</p><p><strong>google-docs-mcp — full Google Workspace integration for Claude Desktop</strong> — Tom Dörr surfaced a-bonus/google-docs-mcp, an MCP server giving Claude Desktop full Google Workspace access — Docs, Sheets, Drive, the lot. For anyone who lives inside Workspace, this turns Claude Desktop into a first-party drafting and editing surface without round-tripping through ChatGPT plugins or copy/paste. Another data point that the MCP ecosystem keeps filling integration gaps the official Claude apps don&#39;t ship natively. Worth evaluating against the &quot;managed MCP&quot; path Google itself is now pushing (<a href="https://x.com/tom_doerr/status/2054392403793002779">tom_doerr</a>)</p><p><strong>Table Reservation Goat — agent-callable CLI for OpenTable + Tock</strong> — Pejman launched Table Reservation Goat CLI on Printing Press, an agent-callable CLI that searches OpenTable and Tock for restaurant availability with advanced booking. Designed to plug into OpenClaw/Hermes, Claude Desktop, or any coding agent via npx -y @mvanhorn/printing-press install table-reservation-goat. Real-world example of the Printing Press distribution pattern — agent-callable CLIs published as installable npm packages, ready to be picked up by Code Mode-style runtimes. Restaurant booking joins the small but growing list of &quot;agentic consumer tasks&quot; delivered as composable CLI primitives rather than full apps (<a href="https://x.com/pejmanjohn/status/2053500526084727295">pejmanjohn</a>)</p><p><strong>Steno — open-source, local-first meeting stenographer for macOS</strong> — Steno is an open-source macOS app that records, transcribes, summarizes, and lets you query confidential meetings — all locally, with recordings, transcripts, and summaries staying on the Mac. Positioned as the privacy-first alternative to Granola, Fireflies, and Otter for orgs that can’t send meeting audio to a cloud vendor. The bookmark-to-like ratio skews heavily toward bookmarks — audience is clearly saving it for later evaluation. Relevant for any meeting workflow that touches sensitive customer or board content (<a href="https://x.com/githubprojects/status/2054086749421457506">githubprojects</a>)</p><p><strong>Layers — Jamie Mill’s new AI plugin worth adding to your stack</strong> — PM George Nurijanian flags Layers (layers.jamiemill.com) as a strong addition to a personal AI-skills collection, putting it alongside @impeccable_ai. Light on technical detail but useful as a signal from the prodmgmt community: a builder-oriented PM curating AI tools as a daily-driver kit. Worth a look for anyone evaluating the growing “AI skills/plugins for individual workflows” wave (<a href="https://x.com/nurijanian/status/2053393755563270401">nurijanian</a>)</p><h3>Product Management</h3><p><strong>Intercom becomes Fin — incumbents torch the brand to escape the old map</strong> — Eoghan McCabe announced Intercom is renaming the parent company to Fin, after the AI customer-service agent they launched three years ago. Intercom (the help-desk product) keeps its name and just shipped Intercom 2, but all 1,400 employees now work for Fin, and Fin is about to be the largest part of the business. McCabe’s argument: humans rarely update their mental map of incumbents, so newer entrants in the AI agent category win despite “provably inferior technology” simply because they have no baggage. A rare incumbent playbook for pivoting into AI-native categories — destroy the brand equity to escape the old map (<a href="https://x.com/eoghan/status/2054238946612367818">eoghan</a>)</p><p><strong>Aakash Gupta’s “Team OS” — pick the pattern by AI weekly-usage rate</strong> — Aakash extends his “Team OS” thesis (a shared repo where every customer call summary, decision log, and analytics query is checked in, queryable in plain English) into a maturity model keyed to AI adoption rate. Under half your team using AI weekly → <strong>Hub-and-Spoke</strong> (one power user queries, posts synthesized output — Dave at Pendo’s morning portfolio briefs). Over half → <strong>Full Adoption</strong> (everyone in the repo, everyone querying, everyone opening PRs — Hannah Stulberg at DoorDash across designers, engineers, data scientists, strategy). Ten+ people with eng resources → layer <strong>Agent Delegation</strong> on top (Gabor Meyer at Google built 21 specialized agents reviewing every spec from their assigned lens). The constraint sits upstream of the tool — count how many teammates use AI weekly before picking the model (<a href="https://x.com/aakashgupta/status/2053504608564314362">aakashgupta</a>)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=083a751529a3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Context Window — W19 2026]]></title>
            <link>https://medium.com/@smnbss/context-window-w19-2026-d58eecc081de?source=rss-5de037daf82f------2</link>
            <guid isPermaLink="false">https://medium.com/p/d58eecc081de</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Simone Basso]]></dc:creator>
            <pubDate>Mon, 11 May 2026 05:29:10 GMT</pubDate>
            <atom:updated>2026-05-11T05:29:10.738Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>The hyperscalers shipped their entire agent stack in seven days.</strong> <br>Google Cloud rolled out Knowledge Catalog, Agent Studio (visual to ADK), Builders Hub, Agent Gateway, Data Insights, Agent Marketplace, and Gemini Agent Mode in a single week (<a href="https://x.com/googlecloudtech/status/2051013926347751457">googlecloudtech</a>, <a href="https://x.com/googlecloudtech/status/2051738707305042031">googlecloudtech</a>, <a href="https://x.com/wesroth/status/2052357792305156250">wesroth</a>). Anthropic shipped Claude Managed Agents with dreaming, multiagent orchestration, outcomes, and webhooks, plus financial-services templates and OIDC-based keyless auth (<a href="https://x.com/claudeai/status/2052067399088664981">claudeai</a>, <a href="https://x.com/claudeai/status/2051679629488865498">claudeai</a>, <a href="https://x.com/claudedevs/status/2051393709619732758">claudedevs</a>). Microsoft showed up with Satya’s “reconceptualize work” essay — slower, more strategic, less ship-y (<a href="https://x.com/satyanadella/status/2051787232043020719">satyanadella</a>). If you’re a startup still pitching “we have an agent platform,” that window closed this week.</p><p><strong>“Skills” is now the cross-vendor word.</strong> <br>Google shipped Skills inside the Gemini app <em>and</em> the Chrome browser as the new way to save and invoke prompts (<a href="https://x.com/wesroth/status/2052357792305156250">wesroth</a>, <a href="https://x.com/googlechrome/status/2052795698626633745">googlechrome</a>). Richard Seroter from Google publicly endorsed agent skills as “the phases a senior engineer navigates, because shipping without them produces incidents” (<a href="https://x.com/rseroter/status/2052032133724799212">rseroter</a>). Anthropic was already there. The two largest agent platforms now use the same primitive name with the same abstraction — that doesn’t happen by accident. The unit of distribution stopped being apps, then stopped being MCP servers, and now it’s Skills. Plan your internal agent strategy around that.</p><p><strong>Agent-native CLIs are eating MCP-per-service sprawl.</strong> <br>Strukto’s Mirage rewrote bash from scratch — 1.1M LOC in six weeks — so cat, grep, and pipes work across Slack, GDrive, GitHub, Linear, Postgres as one POSIX tree (<a href="https://x.com/zechengzh/status/2052105012172792061">zechengzh</a>). Matt Van Horn shipped Printing Press: a factory plus 30+ agent-native CLIs (Linear, ESPN, Flight GOAT, Contact Goat) all local, SQLite-backed, designed to burn the minimum number of tokens per call (<a href="https://x.com/mvanhorn/status/2052422567181611010">mvanhorn</a>). Peter Steinberger&#39;s gog 0.16 ships lossless raw API output and sanitized Gmail reads for LLM ingestion (<a href="https://x.com/steipete/status/2051575048348074450">steipete</a>). The math is brutal: a five-server MCP stack burns 55K tokens before any work; purpose-built CLIs are an order of magnitude cheaper. MCP isn&#39;t dying — the load-everything-upfront pattern is.</p><p><strong>Brain-tools are now a real category.</strong> <br>Elvis at DAIR.AI runs an LLM Wiki plus interactive HTML artifacts as a bidirectional personal control plane — markdown for durable context, HTML for action (<a href="https://x.com/omarsar0/status/2052850591584383177">omarsar0</a>). agentmemory benchmarked a 92% token reduction over the CLAUDE.md pattern on 240 real sessions, with 200x more tool calls before context bites (<a href="https://x.com/ghumare64/status/2052313902214476192">ghumare64</a>). ClickUp shipped Brain² as a model-agnostic context layer that auto-injects projects, decisions, and preferences for any frontier model (<a href="https://x.com/testingcatalog/status/2052828735720903095">testingcatalog</a>). gbrain v0.31.1 went client-server — one home server, every other machine connects over MCP (<a href="https://x.com/garrytan/status/2053306243704410460">garrytan</a>). Same shape as what this brain repo is building. The category is real, and the productized versions are coming fast.</p><p><strong>The chat window is dead. The interface is back.</strong> <br>Brian Chesky said it straight: chat is the wrong UX for travel and e-commerce — agents will drive richer, more dynamic interfaces, not text-forward chatbots (<a href="https://x.com/tbpn/status/2052904435656077578">tbpn</a>). Elvis is already building it with HTML artifacts on top of LLM Wikis (<a href="https://x.com/omarsar0/status/2052850591584383177">omarsar0</a>). Tobi put Shopify’s “River” agent only in public Slack — never DMs — so the whole org learns by osmosis like apprentices overhearing senior engineers (<a href="https://x.com/tobi/status/2053121182044451016">tobi</a>). Rahul Garg reframed Palantir as ontology + orchestration with LLMs as one component, not an AI product (<a href="https://x.com/rahul_garg/status/2051919689333522583">rahul_garg</a>). The interesting agent UX isn’t a chat box. It’s operational, visual, and embedded in the work surface. Build accordingly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*dqV8G1ziMnthK4XtXkKvZQ.png" /></figure><h3>AI / Agents</h3><p><strong>Claude Managed Agents adds dreaming, multiagent orchestration, outcomes, and webhooks</strong> — Live from Code with Claude: agents now run self-learning loops between tasks (“dreaming”), grade themselves against rubrics (outcomes), spawn and coordinate other agents, and trigger/observe runs via webhooks. Managed Agents went from single-shot tool runner to a continually-improving managed runtime closer to Replit Agent and Devin, but with Claude’s reliability profile underneath (<a href="https://x.com/claudeai/status/2052067399088664981">claudeai</a>, <a href="https://x.com/claudedevs/status/2052069321355182447">claudedevs</a>)</p><p><strong>Anthropic signs compute partnership with SpaceX, raises Claude Code and API limits</strong> — Anthropic announced a compute deal with SpaceX that “substantially” increases capacity, on top of recent deals; immediate user-facing effect is higher Claude Code and API usage limits. The SpaceX angle is unusual — likely a Starlink/edge data-center play — and signals Anthropic is diversifying beyond AWS and Google to keep up with rate-limit complaints. For heavy Claude Code users, the daily ceiling pressure should ease (<a href="https://x.com/claudeai/status/2052060691893227611">claudeai</a>)</p><p><strong>Anthropic ships keyless auth for the Claude Platform</strong> — Developers can now authenticate via browser CLI, or let workloads use their existing cloud identity from AWS, GCP, Azure, or any OIDC token provider. Eliminates the API-key rotation pain that has been one of Claude’s top enterprise security gripes. Brings Claude’s auth model in line with how modern cloud-native apps actually do identity (<a href="https://x.com/claudedevs/status/2051393709619732758">claudedevs</a>)</p><p><strong>Anthropic ships Claude agent templates for financial services</strong> — Ready-to-run templates for pitch building, valuation reviews, and month-end close. Install as plugins in Cowork and Claude Code, or deploy in production via the new Managed Agents cookbook. 8.4M views, 26K likes — same playbook as Microsoft and Google: from horizontal foundation-model APIs to vertical pre-built agent bundles where the workflow shape is the moat (<a href="https://x.com/claudeai/status/2051679629488865498">claudeai</a>)</p><p><strong>Mirage — unified POSIX filesystem for agents across SaaS tools</strong> — Open-source virtual filesystem that mounts S3, Google Drive, Slack, Gmail, GitHub, Linear, Notion, Postgres, MongoDB, SSH as one filesystem. Bash rewritten from scratch (1.1M LOC, six weeks) so cat, grep, head, pipes all work across services. ~497K views in a day, 4.6K bookmarks. Directly targets the &quot;MCP per service&quot; sprawl slowing down agent platforms (<a href="https://x.com/zechengzh/status/2052105012172792061">zechengzh</a>)</p><p><strong>Printing Press — a factory and library of agent-native CLIs</strong> — One CLI factory plus 30+ ready-made agent-native CLIs (Linear, ESPN, Flight GOAT for nonstop flights, Contact Goat for LinkedIn). The /printing-press command generates new ones for any service. All local, SQLite-backed, fast — work in Claude Code, Codex, OpenClaw, and Hermes. Thesis: most APIs, MCPs, and official CLIs waste agent tokens; purpose-built tools tighten the loop (<a href="https://x.com/mvanhorn/status/2052422567181611010">mvanhorn</a>)</p><p><strong>Google Cloud Agent Gateway — air traffic control for agent fleets</strong> — Control plane inside the Gemini Enterprise Agent Platform that lets orgs manage entire agent fleets from one point: secure agent-to-tool connectivity, centralized policy enforcement. Hyperscaler answer to the same problem MCP brokers and agent identity startups are racing to solve. For any team running more than a handful of agents, the question of where this control plane lives is now a foundational architectural decision (<a href="https://x.com/googlecloudtech/status/2051738707305042031">googlecloudtech</a>)</p><p><strong>Google launches Agent Mode in Gemini with Skills and Scheduled Actions</strong> — Gemini moves from conversational chatbot to proactive autonomous assistant. Skills = reusable procedural recipes for multi-step tasks. Scheduled Actions = run skills automatically on cadence or trigger. Aligns Gemini’s surface directly with Claude’s Skills primitive — “skill + schedule” is now the cross-vendor consumer-agent standard (<a href="https://x.com/wesroth/status/2052357792305156250">wesroth</a>)</p><p><strong>Google Cloud adds Agent Marketplace to Gemini Enterprise’s Agent Gallery</strong> — Unified destination for agents built by Google, internally by customers, or sourced from partners — discovery, governance, and partner monetization in one surface. Mirrors Salesforce’s AgentExchange and Microsoft’s Copilot Studio catalog. The “agent app store” is now table stakes for every enterprise platform (<a href="https://x.com/googlecloud/status/2052463476954599794">googlecloud</a>)</p><p><strong>Google Cloud Knowledge Catalog — context primitive for agents</strong> — Build business context once, let agents reference it for consistent, rule-aware behavior instead of guessing at business rules. 615 likes, 46.7K views — one of the most engaged GCP posts of the week. Signal: Google wants enterprise knowledge as a first-class primitive in agentic architectures, not a RAG afterthought (<a href="https://x.com/googlecloudtech/status/2051013926347751457">googlecloudtech</a>)</p><p><strong>Gemini Enterprise gets Data Insights agent out of the box</strong> — Agent ships preinstalled, capable of joining structured database data with unstructured sources for cross-cutting operational visibility — no custom development. The pitch: AI-driven analytics from an agent, not a dashboard. Lowers the barrier for enterprises that want answers conversationally rather than building yet another BI surface (<a href="https://x.com/googlecloud/status/2051331018380968176">googlecloud</a>)</p><p><strong>Agent building lands in Gemini Agent Studio (visual → ADK)</strong> — Build agents visually from simple prompts, then export to ADK (Agent Development Kit) for full-code customization. Low-code on-ramp that doesn’t dead-end — developers can graduate to code when they outgrow the visual builder. The right shape for enterprise adoption (<a href="https://x.com/googlecloudtech/status/2050968627914965442">googlecloudtech</a>)</p><p><strong>ClickUp Brain² — model-agnostic context layer over org data</strong> — Auto-injects projects, org structure, and decisions as context for any frontier model (Claude 4.7, GPT 5.5, Gemini 3.1) the user picks. Compresses team knowledge before the prompt reaches the model, retains role/team/format preferences across sessions, integrates Gmail, GitHub, Figma, and Slack natively via MCP. Plus Brain Slides for one-prompt deck generation. Productized version of what we’re stitching together with gbrain + service .AGENT.MD files: vendor-supplied context layer, BYO model (<a href="https://x.com/testingcatalog/status/2052828735720903095">testingcatalog</a>)</p><p><strong>LLM Wikis + HTML Artifacts as a personal-agent operating layer</strong> — Elvis (DAIR.AI) describes his daily stack: an LLM Wiki for durable context plus interactive HTML artifacts that call into agents and get called back. Yields zero-inbox triage, topic monitoring, fast prototyping, deep research, experiment design, scheduled research. Markdown is the substrate; HTML is the dynamic surface. Power users are converging on “context store + dynamic UI + agent” — and it looks a lot like a personal control plane (<a href="https://x.com/omarsar0/status/2052850591584383177">omarsar0</a>)</p><p><strong>agentmemory — long-lived, compressed memory for Claude Code, Codex, Hermes</strong> — Open-source memory layer (2.3K stars, EPL-2.0) that records what coding agents do across sessions, AI-compresses the trace, and re-injects only the relevant slice. At 240 observations, CLAUDE.md costs ~22K tokens; agentmemory does the same job in ~1.9K — a <strong>92% reduction</strong>. Benchmarked on 240 real coding sessions: up to 95% fewer tokens per session and 200x more tool calls before context limits bite (<a href="https://x.com/ghumare64/status/2052313902214476192">ghumare64</a>)</p><p><strong>gbrain v0.31.1 ships MCP thin-client — one home server, many clients</strong> — Real client-server support over MCP: run one home gbrain server, every other machine connects with essentially-local performance. Direct relevance to this brain — we’re in local-stdio + PGLite (single-writer lock issues); thin-client MCP lets a phone, laptop, and remote dev box all share one PGLite/Postgres-backed brain without re-importing or fighting locks (<a href="https://x.com/garrytan/status/2053306243704410460">garrytan</a>)</p><p><strong>GBrain to support multi-topology — local, thin-client, hybrid</strong> — Earlier in the week Garry teased the same direction: local PGLite for solo devs, thin clients for teams, hybrid for centralized planning with local code privacy. Echoes the federation pattern Mintlify, Cursor, and Continue are converging on for code-aware agent context (<a href="https://x.com/garrytan/status/2052394084850159947">garrytan</a>)</p><p><strong>GStack v1.26.3 integrates with GBrain for richer coding-agent context</strong> — Coding agents now get significantly more context: GStack taps GBrain’s knowledge base while writing code, reducing hallucinations. Part of the broader trend of pairing coding assistants with persistent memory/knowledge layers — Cursor, Continue, and now GStack all heading the same way (<a href="https://x.com/garrytan/status/2051339045410885726">garrytan</a>)</p><p><strong>Satya Nadella frames agentic systems as a work-reinvention mandate</strong> — Microsoft CEO argues every firm will need to “reconceptualize work” as it builds agentic systems — AI as an expansion of human agency through org redesign, not just tooling. Links to Microsoft’s 2026 Work Trend Index on agents and human agency. The message lands at a moment when most enterprises are wiring agents into existing org charts instead of redesigning the work — Microsoft wants the narrative to be about org design, not tooling (<a href="https://x.com/satyanadella/status/2051787232043020719">satyanadella</a>)</p><p><strong>Shopify’s “River” agent runs in public Slack only — apprenticeship for the org</strong> — Tobi’s design choice: River never operates in DMs. The argument: visible AI collaboration creates organizational learning by osmosis, like apprentices overhearing senior engineers. He invokes the German <em>Lehrwerkstatt</em> — the agent isn’t replacing the apprentice; it’s creating a shared learning surface for the whole company. Strong signal: where you put your agents shapes whether their knowledge compounds or evaporates (<a href="https://x.com/tobi/status/2053121182044451016">tobi</a>)</p><p><strong>Zapier opens early access to its agent-based automation product</strong> — For AI-forward teams wanting collaborative automation through agents instead of chat interfaces. Used internally for months — 1M+ actions already executed. Positions Zapier as a serious contender in agentic automation alongside LangChain and CrewAI, but with the distribution moat of being already-in-every-stack (<a href="https://x.com/wadefoster/status/2051334727974469931">wadefoster</a>)</p><p><strong>Ashpreet Bedi: how to build an agent platform with Agno + Railway</strong> — End-to-end guide: runtime, storage, connectors, interfaces, infra. Docker local dev + FastAPI + Postgres, deployed to Railway for ~$20/month, JWT auth, Slack/Discord integrations. Comes with os.agno.com reference and a Railway template. 793 bookmarks vs 375 likes — high bookmark/like ratio means devs are saving it as a reference architecture (<a href="https://x.com/ashpreetbedi/status/2052413981487427871">ashpreetbedi</a>)</p><p><strong>Anthropic drops free workshop on architecting multi-agent business operations</strong> — Previously delivered as enterprise briefings, now public. Teaches how to architect, build, and deploy AI agents that run entire business functions autonomously — research, content, customer comms, ops, analytics — not just augmenting tasks. Pitch: 3–5 person teams operating at the level of 50-person orgs. Bookmark/like ratio ~2.8x. The gap between companies treating AI as chat and companies architecting multi-agent systems is widening monthly (<a href="https://x.com/cyrilxbt/status/2052400902745555344">cyrilXBT</a>)</p><p><strong>TestingCatalog: Gemini Agent tasks lean on connectors and research</strong> — Early-look list of Gemini Agent task recipes: Declutter your inbox, Prepare for the meeting, News digest, Bill tracker, Ghost writer. Most lean on connectors and research workflows, not IDE tool use. Read: Google is positioning Gemini Agent as a personal-productivity assistant first, leaving the dev-tools agent space to Claude/Codex — a deliberate market split worth watching (<a href="https://x.com/testingcatalog/status/2052089755483955424">testingcatalog</a>)</p><p><strong>Richard Seroter: agent skills are the engineering phases a senior engineer navigates</strong> — Google Cloud’s Seroter endorses Addy Osmani’s framing: “Skills push the agent through phases a senior engineer navigates, because shipping without them produces incidents.” Strong endorsement from the Google side that skill-driven workflows aren’t a Claude thing — they’re the discipline that separates demo-grade agents from production-shippable ones (<a href="https://x.com/rseroter/status/2052032133724799212">rseroter</a>)</p><p><strong>“This is what Palantir is” — agentic AI as ontology + orchestration</strong> — Rahul Garg quote-tweets a layered mental model for agent design (perception, planning, memory, tool use, orchestration) with the line “This is what Palantir is.” The takeaway: enterprise agent platforms aren’t AI products. They’re ontology + orchestration platforms with LLMs as one component. Right framing for any architecture conversation about an internal agent stack (<a href="https://x.com/rahul_garg/status/2051919689333522583">rahul_garg</a>)</p><p><strong>MCP livestream: implementing Model Context Protocol with AlloyDB and BigQuery</strong> — Google Cloud Tech announced a livestream on implementing MCP for databases — AlloyDB and BigQuery as first-class targets. Notable because MCP is spreading from developer tools into enterprise database infrastructure, which significantly expands the protocol’s footprint (<a href="https://x.com/googlecloudtech/status/2051376314855084529">googlecloudtech</a>)</p><h3>Engineering / DevOps</h3><p><strong>Chrome 146 introduces WebMCP Early Preview for structured agent access</strong> — Developers can expose structured data to AI agents as tools instead of forcing scraping or excessive token consumption. Agents interact with web apps through a defined protocol instead of parsing HTML — fundamentally changing how agents access web data. Major step toward making the web natively agent-friendly and removing the fragility of scraping-based approaches (<a href="https://x.com/chromiumdev/status/2051328272185581963">chromiumdev</a>)</p><p><strong>Chrome DevTools for agents adds viewport and user-agent emulation</strong> — AI agents can now mimic real device conditions when automating or testing sites — closes a long-standing gap where agents driving Chrome via CDP had to fake mobile context manually. Strong unlock for QA/test agents and SEO scrapers needing real-device-rendering parity. Google is steadily turning Chrome into the canonical agent browser runtime (<a href="https://x.com/chromiumdev/status/2052397565904040312">chromiumdev</a>)</p><p><strong>Chrome ships Skills GA — saved customizable prompts in the browser</strong> — Generalizes the saved-prompt / slash-command pattern (Claude Code skills, ChatGPT custom GPTs) to the default browser surface. Meaningful because it lowers activation energy for AI workflows for the long-tail user who never opens a terminal or a chat app. “Skills in your browser” is the consumer entry point (<a href="https://x.com/googlechrome/status/2052795698626633745">googlechrome</a>)</p><p><strong>gog 0.16 — Google Workspace CLI for humans and agents</strong> — Peter Steinberger ships v0.16: lossless raw API output (no lossy reformatting), sanitized Gmail reads (HTML/MIME stripped for LLM ingestion), safer command profiles, Drive inventory, Docs tabs, Sheets tables, Gmail filter export, official Docker images. Exactly the agent-grade CLI shape this brain already uses with gws — gog is a strong alternative when you want lossless raw payloads and containerized runs (<a href="https://x.com/steipete/status/2051575048348074450">steipete</a>)</p><p><strong>Huly — open-source self-hosted bundle replacing Linear + Jira + Slack + Notion</strong> — 25.5K stars, EPL-2.0, self-hosted. Bundles Linear-grade issue tracking + Jira-style PM + Slack-style chat + Notion-style docs + virtual office rooms + GitHub sync + HRMS/ATS/CRM, all in one Docker compose. Pitch math: a 50-person team running Linear Business + Jira Standard + Slack Business+ + Notion Business pays ~$35,490/yr; Huly is free with unlimited users. Worth tracking even without adopting — the “one connected platform” thesis (task ↔ chat ↔ doc ↔ call) is the right direction (<a href="https://x.com/heynavtoor/status/2053026329520673203">heynavtoor</a>)</p><p><strong>Google Cloud Builders Hub — single pane for Cloud, Firebase, and AI Studio</strong> — One destination unifying Google Cloud, Firebase, and AI Studio projects/apps — addresses the long-standing pain of context-switching across three separate consoles. Plus tailored learning-path recommendations. Quiet but overdue DX consolidation: Google is finally treating its three developer surfaces as one product instead of three lineages (<a href="https://x.com/googlecloudtech/status/2051784002290266192">googlecloudtech</a>)</p><h3>Product Management</h3><p><strong>Brian Chesky: chat is the wrong interface for travel and e-commerce</strong> — Airbnb’s CEO argues the post-app future is agentic but visual, not chat-first: “the future is agents, but I don’t think they’re going to be text-forward — they’re going to be really rich user interfaces.” He compares chat-everything to “imagine using iMessage to do everything.” Direct read for WeRoad: agentic doesn’t mean we throw away the rich travel UI — it means the agent drives a richer, more dynamic UI (<a href="https://x.com/tbpn/status/2052904435656077578">tbpn</a>)</p><p><strong>Google Health replaces Fitbit, imports from any tracker</strong> — New health &amp; wellness suite replacing the Fitbit app entirely. Imports from any tracker (not just Fitbit hardware) and embeds Gemini as an AI coach. Google retiring the Fitbit brand at the app layer and consolidating quantified-self under a Gemini-first product. Mirrors Apple Health’s positioning but with a built-in LLM advisor on top (<a href="https://x.com/mweinbach/status/2052389818798104997">mweinbach</a>)</p><p><strong>Luca Rossi: how he runs Tolaria as a one-person product</strong> — Tolaria release with QoL fixes (configurable note width, sidebar pluralization toggle, better date picker, richer whiteboard tools, neighborhood mode, embeds). More interestingly: a long Refactoring piece explaining the full set of workflows he uses to run Tolaria solo. Useful read for anyone running a small product team — concrete view into how a senior PM actually structures their week (<a href="https://x.com/lucaronin/status/2052033007154827512">lucaronin</a>)</p><h3>Design</h3><p><strong>Squidler: 97% of homepages have accessibility issues</strong> — Continuous a11y scanning tool with the pitch that the difference isn’t whether you have issues — it’s whether you find them before users do. Free to start. Category is becoming increasingly relevant as WCAG enforcement tightens in EU and US public sectors. Low-friction baseline check worth a look (<a href="https://x.com/squidlerio/status/2051276003025019142">squidlerio</a>)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d58eecc081de" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>