<?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 Dorian Smiley on Medium]]></title>
        <description><![CDATA[Stories by Dorian Smiley on Medium]]></description>
        <link>https://medium.com/@dorians?source=rss-a10153e4835a------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/2*RtUqOeQI3le_2iHnO62VLA.jpeg</url>
            <title>Stories by Dorian Smiley on Medium</title>
            <link>https://medium.com/@dorians?source=rss-a10153e4835a------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 07:39:02 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@dorians/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[How Ontology Value Compounds: From Controlled Chaos to Enterprise Coherence]]></title>
            <link>https://medium.com/codestrap/how-ontology-value-compounds-from-controlled-chaos-to-enterprise-coherence-b114bc255da6?source=rss-a10153e4835a------2</link>
            <guid isPermaLink="false">https://medium.com/p/b114bc255da6</guid>
            <category><![CDATA[ontology]]></category>
            <category><![CDATA[palantir]]></category>
            <dc:creator><![CDATA[Dorian Smiley]]></dc:creator>
            <pubDate>Thu, 12 Nov 2026 23:04:17 GMT</pubDate>
            <atom:updated>2026-11-12T23:04:43.220Z</atom:updated>
            <content:encoded><![CDATA[<p>Why the most valuable enterprise ontologies are grown, not designed</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PlMe6VpP7UIHzHWZtDZv8A.png" /></figure><p>Enterprise ontologies create value along two dimensions: <strong>vertically, by deepening the capabilities of individual business domains, and horizontally, by connecting those domains so that value compounds across the enterprise.</strong></p><p>The challenge is that these dimensions evolve at different speeds. Finance, manufacturing, demand planning, and transportation have different priorities, different data, and different definitions of the business. Each domain must be free to solve its own problems while contributing to a larger, connected understanding of the enterprise.</p><p>This creates a fundamental tension: too much centralized control suppresses innovation, while too little produces disconnected systems that cannot share meaning. The solution is neither a perfectly designed top-down enterprise model nor an uncontrolled collection of departmental data products. It is <strong>a well governed enterprise ontology that allows local innovation to flourish while progressively establishing shared meaning across the business. </strong>Think of it as controlled chaos, deliberately evolving toward coherence.</p><h3>Two Dimensions of Ontology Value</h3><p>Consider an enterprise with four domains: Finance, Manufacturing, Demand Planning, and Transportation. Each domain can create value independently. But their greatest potential emerges when their capabilities become connected.</p><h4>Vertical value: Deepening domain intelligence</h4><p>Vertical value comes from increasingly sophisticated representations of a particular business domain. A manufacturing team might begin with a simple ontology describing factories, production lines, products, and production orders.</p><p>That initial model, which we can call v0, provides immediate value. Applications can understand the relationships between production orders, equipment, and facilities without repeatedly reconstructing them from underlying data sources.As the domain matures, the ontology evolves.</p><p>At v1, manufacturing might introduce production capacity, material dependencies, and equipment availability. At v2, it might incorporate maintenance schedules, yield optimization, and production simulations. Each evolution enables more sophisticated applications and decisions.</p><p>The important point is that <strong>value is created long before the domain ontology is complete. </strong>A domain does not need to model everything about manufacturing to deliver something useful. It needs to model enough to solve a valuable problem, then expand as additional use cases justify the investment. This is how an ontology becomes an increasingly valuable operational representation of a business.</p><h4>Horizontal value: Connecting domain intelligence</h4><p>Horizontal value emerges when these independently developed models become connected. Suppose Finance has developed an ontology describing product margins, costs, and profitability. Manufacturing has modeled production capacity. Demand Planning understands forecasted demand, while Transportation understands shipping capacity and delivery constraints.</p><p>Individually, these domains can answer useful questions. Together, they can answer fundamentally more valuable questions:</p><ul><li>Which production orders should we prioritize to maximize profitability?</li><li>How would a supply disruption affect customer commitments and revenue?</li><li>Which changes to our demand forecast have the greatest effect on working capital?</li></ul><p>These questions cannot be answered adequately within any single domain. They require relationships spanning multiple domains, with enough shared semantic understanding to interpret the resulting information correctly.</p><p>For example, connecting Finance’s understanding of product profitability to Manufacturing’s production orders makes it possible to optimize production against financial outcomes. Connecting those same production orders to Demand Planning enables decisions that consider both profitability and anticipated demand. Connecting Transportation adds the ability to evaluate whether those plans can actually be fulfilled. <strong>Every new relationship can increase the value of capabilities that already exist.</strong></p><p>This is where enterprise ontology value begins to compound. A relationship established today can be reused by applications, analytics, workflows, and AI agents that did not exist when the object was created. And because domains evolve independently, a relatively mature domain can immediately create value for a much less mature one.</p><p><strong>Horizontal interoperability does not require uniformity, but it does require semantic meaning.</strong></p><h3>Why Purely Top-Down Modeling Fails</h3><p>The traditional temptation is to design the entire enterprise ontology before allowing individual domains to build upon it. A centralized architecture team defines the entities, relationships, taxonomies, and standards that every department must adopt.</p><p>Conceptually, this seems reasonable. If everyone agrees on a common model before building anything, interoperability should follow naturally. In practice, this approach encounters several problems.</p><ol><li>First, enterprise knowledge is distributed. The people who understand manufacturing constraints are not necessarily the people who understand financial accounting, demand forecasting, or transportation optimization.</li><li>Second, business requirements constantly evolve. A model designed centrally today may fail to capture an important operational distinction discovered tomorrow.</li><li>Third, a comprehensive enterprise model requires significant investment before its value can be extracted. Often the window of opportunity shuts before the model is complete.</li></ol><p>The result is often an architecture program that produces elaborate models but struggles to deliver operational outcomes. Traditional Master Data Management (MDM) addresses an important part of this problem by establishing reliable identities, reference data, and authoritative records for key business entities. But MDM alone does not solve the broader challenge of expressing business meaning or extracting business value within an accessible timeframe.</p><h3>Controlled Chaos: Let the Business Lead</h3><p>An alternative is to allow business domains to develop their own ontologies in response to immediate needs. Manufacturing models the production processes that matter to manufacturing. Finance models the financial concepts that matter to Finance. The organization does not require every team to agree on a comprehensive enterprise model before creating value.</p><p>This inevitably produces some duplication, semantic disagreements, and inconsistencies. That is not necessarily failure. It is often a natural consequence of discovering how the business actually operates. Two departments may both define a concept called <em>Revenue</em>, yet mean different things. Finance may care about recognized revenue under accounting rules. Sales may care about bookings. Operations may care about fulfilled orders and their associated economic value. Forcing these concepts into a single definition would eliminate distinctions that matter.</p><p>A well-designed ontology can preserve these domain specific meanings while defining explicit relationships among them. This is the difference between uncontrolled fragmentation and controlled semantic diversity. The objective is to make those differences understandable, governed, and interoperable.</p><h3>Governance Makes Organic Growth Possible</h3><p>Controlled chaos requires governance. Without governance, independently developed ontologies risk becoming another generation of disconnected data silos. The trick is to governance that enables interoperability without prescribing every detail of how domains model their businesses.</p><p>This requires establishing a small but essential set of enterprise-wide rules:</p><ol><li><strong>Clear ownership.</strong> Every domain owns its concepts, definitions, relationships, and quality expectations. Cross-domain relationships have explicit stewardship.</li><li><strong>Semantic contracts.</strong> Published concepts and relationships have defined meanings, types, constraints, and compatibility expectations that other domains can safely depend upon.</li><li><strong>Versioning and evolution.</strong> Domains can extend their models without silently breaking the applications and relationships that depend upon them.</li><li><strong>Policy and provenance.</strong> Access controls, lineage, and data quality remain enforceable as information crosses domain boundaries.</li><li><strong>Federated governance.</strong> Central governance establishes interoperability standards, while domain experts retain authority over their business semantics.</li></ol><p>These guardrails create a framework within which teams can innovate independently. A concept used by one application within a single domain might require relatively little coordination. A concept depended upon by dozens of applications across Finance, Manufacturing, and Operations demands stronger semantic guarantees. As relationships become more valuable, the organization invests more heavily in governing them. Governance evolves alongside adoption rather than becoming a prerequisite for it.</p><h3>The Compounding Effect</h3><p>The economic significance of this approach becomes clearer over time. Imagine Finance is the first domain to establish a useful ontology. Its initial value is limited to financial applications and decisions. Manufacturing subsequently develops its own ontology and connects production orders to financial measures. Both domains now become more useful. Demand Planning then connects forecasts to production orders, product profitability, and inventory. Transportation introduces fulfillment constraints, shipping costs, and delivery commitments. Each new domain contributes its own capabilities, but it can also reuse the semantic foundation established by the others.</p><p>The enterprise is no longer simply accumulating isolated applications. It is accumulating reusable business understanding. A new application addressing product profitability no longer needs to independently discover how products relate to manufacturing orders, demand forecasts, and transportation costs. Those relationships are already represented and governed. An AI agent investigating a potential supply chain disruption can traverse these same relationships to understand downstream consequences. The marginal cost of developing new capabilities can decline because an increasing portion of the business context and actions already exists.</p><p>This is the real compounding opportunity: <strong>each investment in use cases creates a broader asset that can increase the value of future investments. </strong>The rate of compounding depends on how composable the use case is from existing assets, how reliably domains can interoperate, and whether teams actually build upon the shared business understanding captured in the enterprise ontology.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b114bc255da6" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codestrap/how-ontology-value-compounds-from-controlled-chaos-to-enterprise-coherence-b114bc255da6">How Ontology Value Compounds: From Controlled Chaos to Enterprise Coherence</a> was originally published in <a href="https://medium.com/codestrap">Codestrap</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Rethinking AI Coding: A Three-Layer Architecture for Software Development]]></title>
            <link>https://medium.com/codestrap/rethinking-ai-coding-a-three-layer-architecture-for-software-development-f211408be7cd?source=rss-a10153e4835a------2</link>
            <guid isPermaLink="false">https://medium.com/p/f211408be7cd</guid>
            <category><![CDATA[ai-coding]]></category>
            <dc:creator><![CDATA[Dorian Smiley]]></dc:creator>
            <pubDate>Wed, 21 Oct 2026 18:49:29 GMT</pubDate>
            <atom:updated>2026-10-21T18:49:48.268Z</atom:updated>
            <content:encoded><![CDATA[<p>The next major improvement in AI coding might come less from better models and more from better software architectures around those models.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gbkcHc9IwWDghDmjXboqCw.png" /></figure><p>At CodeStrap, we’re exploring an architecture built around three complementary layers designed to improve correctness, reduce generation costs, and make AI generated software easier to maintain and evolve.</p><p><strong>Layer 1: Validation Driven Development (VDD)</strong></p><p>The idea is simple: instead of relying on a model to determine whether its own output is correct, developers define explicit validation procedures using CLI scripts, executable rules, and prompts. These procedures establish the invariants that generated software must satisfy.</p><p>These are not tests! Let’s be honest, you are having models write the tests anyway. What we define is a process for validation along with cli scripts to run to verify correctness. The process is inherently nondeterministic, but that is the point. We are trading the determinism of a test for massive surface area. Below is a sample regression test we use to test regressions in our semantic and scaffolding agents.</p><pre># Archie/Factory regression test: BRANCH vs BASE<br><br>```text<br>BRANCH = your-feature-branch<br>BASE   = origin/master<br>```<br><br>Goal: check that BRANCH does not regress the Factory Orchestrator Broker (Factory) or Archie Broker (Archie) compared with BASE. Run Factory and then Archie on both, using the sample application scenario, and compare. Do not change source code or commit anything; report findings with evidence.<br><br>## Setup (from your main checkout; never run generators in the main checkout itself)<br><br>```bash<br>git fetch origin<br>git worktree add --detach .codex/worktrees/reg-base   BASE<br>git worktree add --detach .codex/worktrees/reg-branch BRANCH<br>```<br><br>For each worktree `&lt;WT&gt;`:<br><br>```bash<br>cp .env.local &lt;WT&gt;/.env.local<br>cp apps/cli-tools/.env &lt;WT&gt;/apps/cli-tools/.env<br>cd &lt;WT&gt; &amp;&amp; pnpm install --frozen-lockfile --prefer-offline &amp;&amp; mkdir -p .scratch<br>```<br><br>Never print env files; `SQL_LITE_DB_PATH` must be absolute.<br><br>If a worktree path already exists, remove it first (`git worktree remove --force &lt;path&gt;`) or choose a new name.<br><br>Run base and branch in parallel, since the worktrees are independent. Factory takes about 10–15 min and Archie about 30–40 min.<br><br>## Factory (in each `&lt;WT&gt;`)<br><br>Write `&lt;WT&gt;/.scratch/request.md`, replacing `&lt;WT&gt;` with the worktree&#39;s ABSOLUTE path (`--request-file` does not expand placeholders):<br><br>```text<br>I want to build a new feature in apps/your-app<br>based on the examples/your-example using<br>your application&#39;s SDK.<br><br>Generate a software factory manifest from the<br>workflow screenshots and README in<br>examples/your-example.<br><br>Use &lt;WT&gt;/apps/your-app/src/components as the reusable<br>components root,<br>&lt;WT&gt;/apps/your-app/src/modules as the modules root,<br>&lt;WT&gt;/packages/services/your-provider/src/lib/doa/your-app/delegates<br>as the DAO delegates root,<br>&lt;WT&gt;/packages/services/your-provider/src/lib/doa/your-app<br>as the DAO root,<br>&lt;WT&gt;/apps/your-app/src/app/api as the API routes root,<br>&lt;WT&gt;/apps/your-app/src/hooks as the service hooks root<br>and &lt;WT&gt;/apps/your-app/src/app as the pages<br>App Router root.<br><br>Package name is @your-app/sdk.<br>These paths are confirmed.<br>```<br><br>Then run:<br><br>```bash<br>REPO_ROOT=$PWD WORKSPACE_ROOT=$PWD BASE_DIRECTORY=$PWD pnpm factory:dag:auto -- --request-file .scratch/request.md --input-dir examples/your-example --max-turns 30 --confirm-inputs &gt; .scratch/factory.log 2&gt;&amp;1<br>```<br><br>The last line is `dagPath: &lt;path&gt;`. If the exit code is not 0 or there is no dagPath, retry (up to 3 times).<br><br>If the CLI flags or output differ on either side, check:<br><br>```bash<br>pnpm factory:dag:auto -- --help<br>```<br><br>Adapt as necessary.<br><br>## Archie (in each `&lt;WT&gt;`)<br><br>```bash<br>REPO_ROOT=$PWD WORKSPACE_ROOT=$PWD BASE_DIRECTORY=$PWD pnpm archie:dag:auto -- --dag &lt;dagPath&gt; &gt; .scratch/archie.log 2&gt;&amp;1<br>```<br><br>The execution id is in the newest `.your-runtime/artifacts/archie-dag-auto-&lt;executionId&gt;.json`.<br><br>If Archie exits 1, read the last &quot;runner error&quot;/&quot;stalled&quot; line in the log:<br><br>- `&quot;stalled: ... does not require review and is not terminal&quot;` is a known CLI quirk. Just continue:<br><br>```bash<br>REPO_ROOT=$PWD WORKSPACE_ROOT=$PWD BASE_DIRECTORY=$PWD pnpm archie:dag:auto -- --execution-id &lt;executionId&gt;<br>```<br><br>- Any failed state (e.g. a transient &quot;OpenAI API error 4xx/5xx&quot;): retry that state, then continue as above, up to 6 times.<br><br>Retry helper, saved as `&lt;WT&gt;/.scratch/retryArchie.ts`:<br><br>```typescript<br>import { getYourContainer } from &#39;@your-org/developer-foundations-x-reason&#39;;<br>import { withContainer } from &#39;@your-org/server-utils&#39;;<br>import { yourDagRetryFunction } from &#39;@your-org/brokers&#39;;<br><br>withContainer(getYourContainer(), () =&gt;<br>  yourDagRetryFunction(<br>    process.argv[2],<br>    &#39;{}&#39;,<br>    undefined,<br>    undefined,<br>    true<br>  ),<br>)<br>  .then((r) =&gt; {<br>    console.log(JSON.stringify(r));<br>    process.exit(r.error ? 1 : 0);<br>  })<br>  .catch((e) =&gt; {<br>    console.error(e);<br>    process.exit(1);<br>  });<br>```<br><br>Run it with:<br><br>```bash<br>REPO_ROOT=$PWD WORKSPACE_ROOT=$PWD BASE_DIRECTORY=$PWD DOTENV_CONFIG_PATH=apps/cli-tools/.env NODE_OPTIONS=--max-old-space-size=8192 SENTRY_ENABLED=false TS_NODE_TRANSPILE_ONLY=true node_modules/.bin/ts-node --project apps/cli-tools/tsconfig.cli.json -r dotenv/config -r tsconfig-paths/register -r ./tools/loaders/raw-text-register.cjs .scratch/retryArchie.ts &lt;executionId&gt;<br>```<br><br>If `yourDagRetryFunction`&#39;s signature differs on a side (check `packages/brokers/src/agents/yourDagRetryFunction.ts`), adapt the call so it retries the failed state.<br><br>- The same failure repeating on both sides after 6 attempts is a shared stop, not a branch regression. Record where each side stopped.<br><br>## Compare BASE vs BRANCH<br><br>First read:<br><br>```bash<br>git diff --stat BASE...BRANCH<br>```<br><br>See what BRANCH changes, and look hardest at the outputs those areas produce.<br><br>### Factory<br><br>- Exit code and number of tries; DAG errors/warnings (expect 0).<br>- Node counts per kind and disposition (new/reuse).<br>- Requirements contract: keys, packageName `@your-app/sdk`, generation roots, and any new or missing fields.<br><br>Different node counts alone are model variance; even the same commit varies between runs.<br><br>### Archie<br><br>- Final state (expect success), retries and why.<br>- How often each creator state ran (from &quot;Executing function&quot; lines in the logs, or `context.stack` of the `machine_executions` row in `SQL_LITE_DB_PATH`).<br>- Review steps with GAPS REMAIN and their &quot;### Deterministic gaps&quot; (`.your-runtime/artifacts/batch-review-&lt;executionId&gt;-*.md`).<br>- In each worktree, the files the run created and changed (`git status` / `git diff`): same set of generated artifacts, and no unexpected edits to shared files such as `packages/services/your-provider/src/index.ts` (every generated DAO exported, no delegate exported) or `apps/your-app/.storybook`.<br>- Type errors:<br><br>```bash<br>node node_modules/typescript/bin/tsc6 --noEmit -p apps/your-app/tsconfig.json<br>```<br><br>```bash<br>node node_modules/typescript/bin/tsc6 --noEmit -p packages/services/your-provider/tsconfig.lib.json<br>```<br><br>If BASE and BRANCH differ, rule out Factory variance: run Archie on BRANCH with BASE&#39;s DAG (copy it and replace the base worktree path with the branch worktree path), so both Archie runs share one DAG.<br><br>A difference that persists on the same DAG is evidence of a possible BRANCH regression; account for model variance and environmental differences before attributing it conclusively.<br><br>## Known issues on master<br><br>Confirm these also occur on BASE before dismissing them:<br><br>- `&lt;KNOWN_BASELINE_API_TEST_FAILURE&gt;` — known test execution or repair failure.<br>- `&lt;KNOWN_BASELINE_DEPENDENCY_ISSUE&gt;` — known dependency/import validation gap.<br>- `&lt;KNOWN_BASELINE_TYPE_ERROR&gt;` — known TypeScript validation issue.<br>- The &quot;stalled&quot; CLI quirk.<br><br>## Report<br><br>A BASE vs BRANCH table covering every point above, then a verdict: regression / no regression / inconclusive.<br><br>For each difference give its evidence (file paths, log lines) and say whether it is model variance or attributable to BRANCH.<br><br>Leave the worktrees in place.</pre><p>This prompt is given directly to Claude or Codex to run regression analysis.</p><p>In data engineering, validation might verify that Spark executors are efficiently utilized, data values remain within expected ranges, and performance hasn’t regressed. In application development, validation might enforce REST conventions, API contracts, TypeScript correctness, or compare application behavior against a known baseline.</p><p>The important shift is that correctness becomes something the coding agent must demonstrate through evidence rather than simply assert.</p><p><strong>Layer 2: Scaffolding with Small Models</strong></p><p>Not every line of code needs to be generated by an expensive frontier model. We can use small, specialized models and decompose problems into smaller subproblems to generate much of the initial implementation. The goal is to produce scaffolds that are 80% or more complete at a fraction of the cost.</p><p>Frontier models then take over to complete the remaining implementation, guided by the validation procedures established in Layer 1. This changes the economics of AI coding by reserving expensive reasoning for the portions of development that actually require it.</p><p><strong>Layer 3: Semantic Software</strong></p><p>This is potentially the most important layer for the long-term evolution of AI-generated applications. Today, coding agents spend enormous amounts of time searching repositories, reading files, and trying to infer how components relate to one another.</p><p>At CodeStrap, we’re addressing this with a Merkle graph that represents application components as nodes and their relationships as edges. Interfaces, contracts, and implementations are represented using content-derived hashes, allowing changes to be detected and their potential downstream impact identified.</p><p>Instead of asking a model to grep through thousands of files and guess what might be affected by a change, the semantic layer can provide a much smaller, more precise set of relevant components and dependencies. This has the potential to dramatically reduce context consumption while improving the reliability of changes to large applications.</p><p><strong>Putting It All Together</strong></p><p>The three layers reinforce each other:</p><ol><li>Semantic Software identifies what needs to change and supplies the relevant context.</li><li>Scaffolding agents generate most of the implementation cheaply.</li><li>Validation Driven Development provides the executable evidence needed to establish correctness, with frontier models resolving failures and completing the implementation.</li></ol><p>The result is an architecture designed to make AI coding faster, cheaper, more reliable, and increasingly deterministic. It’s been so successful for us we are able to pay all token costs of our build teams at customers. Full application integrated with Palantir Foundry and Snowflake e2e produce token costs as low as tens of dollars. The token costs are irrelevant for our build teams, so we pay them. That’s an amazing competitive advantage!</p><p>The more we can represent software structure explicitly, generate routine implementations economically, and validate correctness, the less we need to depend on expensive probabilistic reasoning for every development task.</p><p>That’s the direction we’re pursuing at <a href="https://codestrap.com/">CodeStrap</a>. Reach out if you want to learn more.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f211408be7cd" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codestrap/rethinking-ai-coding-a-three-layer-architecture-for-software-development-f211408be7cd">Rethinking AI Coding: A Three-Layer Architecture for Software Development</a> was originally published in <a href="https://medium.com/codestrap">Codestrap</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Death by a Thousand Cuts: AI and the Economics of Big Consulting]]></title>
            <link>https://medium.com/codestrap/death-by-a-thousand-cuts-ai-and-the-economics-of-big-consulting-8d9868316815?source=rss-a10153e4835a------2</link>
            <guid isPermaLink="false">https://medium.com/p/8d9868316815</guid>
            <category><![CDATA[big-4]]></category>
            <category><![CDATA[consulting]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[palantir]]></category>
            <dc:creator><![CDATA[Dorian Smiley]]></dc:creator>
            <pubDate>Thu, 15 Oct 2026 20:26:08 GMT</pubDate>
            <atom:updated>2026-10-15T20:26:32.373Z</atom:updated>
            <content:encoded><![CDATA[<p>The price the market will pay for anything produced by an LLM is going in one direction, and it ain&#39;t up.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xeSM5xBa8KNxhGmEbxvlKQ.png" /></figure><p>If your deliverable is a PowerPoint presentation, the market is going to pay less. If you deliver code, the market is going to pay less. If you produce reports, analysis, documentation, or any other intellectual artifact that AI can increasingly generate, the market is going to pay less.</p><p>When the marginal cost of producing something collapses, competition eventually forces prices down. The question isn&#39;t whether that happens. It&#39;s how quickly businesses can reorganize before the economics of their existing operating models implode.</p><p>The prevailing narrative is that AI will make knowledge workers dramatically more productive. That&#39;s true, but it misses something fundamental. Productivity improvements don&#39;t necessarily translate into increased profits. They can just as easily destroy revenue.</p><p>Imagine a consulting engagement that historically generated $1 million in revenue with $600,000 in delivery costs. That&#39;s $400,000 in gross profit.</p><p>Now imagine AI reduces delivery costs to $100,000. Fantastic, right? Except competitors can do the same thing, and the client knows it. The market reprices the engagement to $300,000. Your gross margin actually improves, but your gross profit just fell by 50%.</p><p>Meanwhile, you&#39;ve got the same partners, management layers, compensation expectations, office space, and overhead built around a very different revenue model. This is the problem with treating AI primarily as a productivity initiative. <strong>You can become dramatically more efficient while your business becomes dramatically less profitable.</strong></p><p>Prices don’t get reset all at once. Pressure is exerted through thousands of individual pricing decisions. Every contract renewal, every competitive bid, every procurement negotiation, every client asking why they&#39;re still paying millions for deliverables that can increasingly be generated in hours.</p><p>A little pricing pressure here. Lower utilization there. A competitor willing to accept half the fees because its operating model supports it. Another client bringing work in-house. Profits die a death of a thousand cuts.</p><p>The structural problem for big consulting is that its economics were built around monetizing scarce human labor. Large teams, billable hours, utilization targets, leverage ratios, and partnership structures all make sense when producing sophisticated intellectual work requires armies of expensive people. As the market discounts the work product, can firms adapt quickly enough to avoid disruption?</p><p>So far, the evidence is largely positive for big consulting. The latest reported results show that the incumbents remain economically resilient.</p><p><a href="https://www.pwc.com/gx/en/about/global-annual-review/financial-performance.html">PwC</a>: FY2025 global revenue of $56.9B, up 2.7% in local currency. Advisory grew 4.5%.</p><p><a href="https://www.deloitte.com/global/en/about/press-room/global-revenue-announcement.html?id=gx:2sm:3li:4GR2019::6abt:20190910113000:connectorsfy20:2626255615&amp;linkId=73447580&amp;nc=1">Deloitte</a>: FY2026 revenue of $74.5B, up 3.8% in local currency.</p><p><a href="https://www.ey.com/en_gl/newsroom/2026/10/ey-global-revenue-increases-4-point-7-percent-to-us-dollars-57b-for-fiscal-year-2026">EY</a>: FY2026 revenue of $57.0B, up 4.7% in local currency. AI service sales grew 49%.</p><p><a href="https://www.sec.gov/Archives/edgar/data/1467373/000146737326000037/q4fy26earnings8-kexhibit.htm?trk=article-ssr-frontend-pulse_little-text-block">Accenture</a>: FY2026 revenue of $74.2B, up 5% in local currency. Consulting revenue grew 3% and managed services grew 6%.</p><p>The Accenture figures are particularly important because operating profits are growing, too.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vCfBga8wwAShcpOG6HVuAg.png" /></figure><p>Similarly, Deloitte UK reported that its FY2026 distributable profits rose 14% on revenue growth of just 2%. That suggests effective cost management can protect partner economics even when revenue growth is modest.</p><p>The counterfactual for what growth could have been is observable with companies like Palantir and the dozens of boutique consulting firms popping up looking to cash in on the FDE model. Forward Deployed Engineering (FDE) is emerging as one of the fastest growing segments of the services industry, with companies like OpenAI and Anthropic committing billions to building out these capabilities.</p><p>How much growth are these firms leaving on the table while they congratulate themselves for improving margins? Palantir’s US commercial revenue grew 149% year over year in Q2 2026. That’s predominantly software revenue, not consulting fees, but the growth is being driven by a fundamentally different approach to enterprise software delivery. One where a software platform accelerates service delivery, and the people delivering those services inform the product roadmap.</p><p>Meanwhile, Deloitte UK’s Technology &amp; Transformation business grew just 4%. These aren’t directly comparable businesses, but that’s precisely the point. The question isn’t whether the incumbents can protect partner profits by cutting costs. It’s whether they’re watching an enormous new market develop around them while they optimize the economics of the old one.</p><p>And Palantir is hardly alone. Forrester documented roughly $9 billion in commitments to forward deployed engineering initiatives from OpenAI, Anthropic, AWS, and Microsoft in just a few months of 2026. OpenAI launched a dedicated deployment company with more than $4 billion in initial investment, while Anthropic backed a $1.5 billion enterprise services venture. Those are investments, not recognized services revenue, but the signal is impossible to ignore. Some of the most sophisticated technology companies in the world are pouring billions into a delivery model built around embedding engineers inside businesses, understanding their operations, and turning that knowledge into working software.</p><p>That sounds an awful lot like consulting, except the economics and organizational structures are fundamentally different. The danger for big consulting is that the next generation of services businesses could be growing exponentially while the incumbents are celebrating single-digit growth and protecting partner distributions. By the time that becomes obvious in their financial statements, the market may have already moved on and disruption could be unavoidable.</p><p>The ultimate winners will be the companies that reorganize around this new economic reality quickly, using software to scale scarce human judgment rather than continuing to sell increasingly commoditized human production. New Co. firms will begin to look less like small versions of the major firms and more like micro-mills: compact, highly automated businesses that take on the work large consultancies are increasingly willing to abandon. Bespoke software. Systems integration. Business-process engineering. The practical implementation work that is often better done by a focused team than by a sprawling organization built around layers of leverage and utilization.</p><p>The large firms will likely lean harder into their strategy groups and traditional management consulting practices. That may work for a while. But the New Co. firms will be busy engineering the business processes and automation systems that actually run the company. They will understand how those systems map to the profit and loss statement. The large consulting firms often will not, and as a result, businesses they used to know and understand will become increasingly opaque.</p><p>A New Co. firm that understands how a business operates at the process level, how automation changes its cost structure, and how those changes flow through the P&amp;L can offer better strategy and advice over time. Its recommendations will be grounded in the machinery of the business rather than separated from it. And that’s what makes this so dangerous for the incumbents.</p><p>Old Co. firms can be extraordinarily well run businesses, optimized for the wrong economic future. They’ll continue optimizing for utilization, protecting partner distributions, and growing revenue while the market slowly reorganizes around an entirely different delivery model. None of those decisions will look catastrophic in isolation. But by the time the aggregate effect becomes obvious, the businesses that understand how to operate in the new economy may already own the client relationships. Death by a thousand cuts.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8d9868316815" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codestrap/death-by-a-thousand-cuts-ai-and-the-economics-of-big-consulting-8d9868316815">Death by a Thousand Cuts: AI and the Economics of Big Consulting</a> was originally published in <a href="https://medium.com/codestrap">Codestrap</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Software Factories Are Platform Engineering. Get Over the Name.]]></title>
            <link>https://medium.com/codestrap/software-factories-are-platform-engineering-get-over-the-name-a14eb7d3a96a?source=rss-a10153e4835a------2</link>
            <guid isPermaLink="false">https://medium.com/p/a14eb7d3a96a</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[platform-engineering]]></category>
            <category><![CDATA[software-factories]]></category>
            <dc:creator><![CDATA[Dorian Smiley]]></dc:creator>
            <pubDate>Tue, 06 Oct 2026 18:24:04 GMT</pubDate>
            <atom:updated>2026-10-06T18:24:31.522Z</atom:updated>
            <content:encoded><![CDATA[<p>AI didn’t invent the economics of scale for software engineering. It just made them accessible to everyone, and increasingly impossible to ignore.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*89kYau5TGDYtrZ6qKbUPRg.png" /></figure><p>I keep hearing engineers ask, “What is a software factory?” It’s platform engineering. That’s it.</p><p>AI has made it economically viable at a scale that was previously inaccessible to most organizations. The principles aren’t new. The engineering practices aren’t new. Some of the best technology companies in the world spent decades figuring this stuff out. But AI has introduced new challenges not the least of which is the rapid accelerated growth of most codebases.</p><h3>Google Figured This Out a Long Time Ago</h3><p>Google was one of the pioneers of engineering at scale. As its engineering organization grew, it had to solve problems that most companies never encountered.</p><p>How do you coordinate thousands of engineers working across millions of lines of code? How do you enforce quality without creating bottlenecks? How do you maintain reliability while continuously deploying changes? How do you prevent every product team from reinventing the same infrastructure?</p><p>You build platforms.</p><p>Google invested heavily in shared build systems, automated testing, code review infrastructure, deployment tooling, and common engineering practices. It solved hard engineering problems once and distributed those solutions across the organization.</p><p>In 2020, Titus Winters, Hyrum Wright, and Tom Manshreck documented many of these lessons in <a href="https://research.google/pubs/software-engineering-at-google/"><em>Software Engineering at Google: Lessons Learned from Programming Over Time</em></a>.</p><p>The book explains the difference between programming and software engineering through three dimensions: time, scale, and tradeoffs. That distinction matters more today than ever. Writing code is not the same thing as engineering software that can be maintained, changed, governed, and operated reliably at scale.</p><p>Google understood that human coordination and manual processes eventually become bottlenecks. Its engineering infrastructure automated enormous amounts of work that would otherwise have required engineers to perform repetitive tasks. That investment helped make it possible to scale engineering while maintaining consistent standards across a massive organization.</p><p>Google wasn’t alone. Other large technology companies reached similar conclusions. What we now call platform engineering emerged from decades of investment in shared infrastructure, developer tooling, DevOps, and operational automation. The lesson was always the same. At scale, you cannot afford to solve the same engineering problem repeatedly and quality must be guaranteed by the platform.</p><h3>Platforms Are the Economies of Scale for Software Engineering</h3><p>A platform is not simply a collection of developer tools. It is an economic model for producing software. Without a platform, every product team independently solves some combination of the same problems: infrastructure provisioning, authentication, security, testing, deployment, observability, and operational reliability. That’s expensive, inconsistent, and difficult to govern. With a platform, those capabilities become shared services.</p><p>You solve hard problems once. You automate quality. You express infrastructure through service catalogs. You standardize provisioning and deployment. You make engineering labor more fungible by reducing the specialized knowledge required to work across products.</p><p>Instead of having ten teams implement ten different deployment pipelines, you build one capability that supports all ten teams. As additional products consume the platform, the cost of developing and maintaining those shared capabilities is amortized across a larger base. That’s how economies of scale work.</p><p>Great platforms allow companies to grow to enormous sizes, maintain attractive margins, and operate with several nines of reliability. Platforms institutionalized knowledge into reusable systems.</p><h3>Why Didn’t Everyone Build Platforms Before?</h3><p>Engineering resources are scarce. Every engineer working on the platform is an engineer who isn’t building a customer facing (revenue generating) product. Product teams can point to features, revenue, and customer demand. Platform teams often have to justify investments based on avoided costs, future productivity, and operational risk.</p><p>Guess who usually wins that argument?</p><p>For decades, companies were forced to choose between investing in their products and investing in the engineering infrastructure that would make future products cheaper to build. Unless you had Google’s scale, it was difficult to justify the fixed investment required to build a truly great platform.</p><p>AI fundamentally changes that equation.</p><p>AI dramatically reduces the labor required to build and maintain many platform capabilities. At the same time, it increases the volume of software those capabilities can support. The cost of establishing the platform comes down while the potential returns increase.</p><p>For the first time, smaller organizations can pursue engineering economies of scale that were historically reserved for companies with enormous engineering budgets.</p><p>There are still architecture decisions, quality requirements, and engineering investments to make. AI doesn’t eliminate judgment. But it changes the resource allocation problem from a zero-sum competition between product and platform teams into an opportunity to invest in both.</p><h3>For the First Time, Everyone Has Scale Problems</h3><p>Gergely Orosz recently published <a href="https://newsletter.pragmaticengineer.com/p/the-state-of-the-tech-industry-in"><em>The State of the Tech Industry in 2026</em></a>. One of the most important observations is how rapidly agentic coding is changing the volume of software being produced.</p><p>Engineers are running five or ten coding agents simultaneously. GitHub reported a ninefold increase in agent-authored pull requests over eight months. Historically, the amount of software an organization could produce was constrained by the number of engineers it employed. That constraint is disappearing.</p><p>A team of ten engineers can now generate a volume of code that would previously have required a much larger organization. But the capacity to generate code is not the same as the capacity to engineer, validate, deploy, and operate software.</p><p>You can’t review ten times as much code with the same manual processes. You can’t govern hundreds of independently generated applications through architecture committees and spreadsheets. You can’t maintain reliability if every agent invents its own infrastructure, security model, and deployment strategy. Everyone suddenly has the engineering scale problems that only companies like Google used to have.</p><h3>DORA Has the Research to Back This Up</h3><p>This isn’t just a theoretical argument.</p><p>Google’s DORA research group examined the relationship between AI adoption and organizational performance in its <a href="https://dora.dev/research/2025/dora-report/">2025 State of AI-assisted Software Development report</a>.</p><p>The central finding is that <strong>AI is an amplifier.</strong></p><p>It amplifies the strengths of well-functioning engineering organizations and the dysfunctions of poorly functioning ones. Simply giving engineers better AI coding tools does not automatically produce better organizational outcomes. In fact, faster code generation can make things worse when downstream processes cannot absorb the increased volume.</p><p>More code means more testing, more security concerns, more integration complexity, more deployment activity, and more operational risk. DORA’s <a href="https://dora.dev/capabilities/platform-engineering/">research on platform engineering</a> is particularly revealing. When platform quality is high, the relationship between AI adoption and organizational performance is strongly positive. When platform quality is low, that relationship is negligible.</p><p>Think about that.</p><p>You can give every engineer an army of coding agents and still fail to improve organizational performance because the surrounding engineering system cannot handle the output. High-performing organizations create standardized, automated pathways for building, testing, securing, and deploying software. AI sits on top of those capabilities and amplifies them. Poorly performing organizations simply automate the production of more problems.</p><p><strong>If you aren’t implementing a platform to support the scale of agentic coding, you are building for collapse.</strong></p><h3>The Platform Is No Longer Optional</h3><p>The software factory extends traditional platform engineering by automating more of the production process itself. Agents can turn specifications into code, assemble applications, generate tests, and execute development workflows. But those agents still need a controlled environment in which to operate.</p><p>They need architectural standards, reusable components, dependency management, automated verification, security policies, deployment infrastructure, telemetry, and mechanisms for managing change. Otherwise, you are just generating code faster than your organization can understand or maintain it.</p><p>This is why I find the obsession with the term <em>software factory</em> somewhat amusing. People called it a software factory because of the throwback to the Industrial Revolution. It makes for nice wallpaper and paintings in the office. But the legacy of the Industrial Revolution wasn’t the art. It was about changing the economics of production through mechanization, standardization, specialization, and economies of scale. It was about turning steam engines into the drivers of economic growth for centuries to come.</p><p>Without the factory the steam engine would have remained a novelty. And its the same today with AI.</p><p>Get over the name, learn the lesson, and start building for scale.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a14eb7d3a96a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codestrap/software-factories-are-platform-engineering-get-over-the-name-a14eb7d3a96a">Software Factories Are Platform Engineering. Get Over the Name.</a> was originally published in <a href="https://medium.com/codestrap">Codestrap</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Agent Loop Is a Recursive Tax]]></title>
            <link>https://medium.com/codestrap/the-agent-loop-is-a-recursive-tax-1e0946271be9?source=rss-a10153e4835a------2</link>
            <guid isPermaLink="false">https://medium.com/p/1e0946271be9</guid>
            <category><![CDATA[loop]]></category>
            <category><![CDATA[tokenomics]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agents]]></category>
            <dc:creator><![CDATA[Dorian Smiley]]></dc:creator>
            <pubDate>Wed, 29 Jul 2026 17:40:59 GMT</pubDate>
            <atom:updated>2026-07-29T21:33:32.425Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*EbkPZS2-WwXCIvzulU9NHA.jpeg" /></figure><p>The default architecture for agentic AI is increasingly a recursive loop with a credit card attached to it. Give the model a goal. Load a massive system prompt. Describe every tool it might possibly use. Inject a pile of organizational context. Ask it what to do. Execute the tool call. Append the result. Send the entire growing conversation back to the model. Ask it what to do again. Repeat until the model decides it is finished, runs out of context, hits a budget limit, or does something stupid enough that a human has to intervene.</p><p>That is the generalized agent loop. It is flexible. It is easy to demonstrate. It can be useful for genuinely open ended work. It is also an extraordinarily inefficient way to run most enterprise processes.</p><p>Anthropic has been unusually honest about its tokenomics. In its <a href="https://www.anthropic.com/engineering/multi-agent-research-system">description of its production research system</a>, Anthropic reported that agents typically consume about four times as many tokens as normal chat interactions. Multi agent systems consume about fifteen times as many. Anthropic also found that token usage alone explained 80 percent of the performance variance on the BrowseComp benchmark. In other words, much of the performance gain came from spending significantly more on inference. This is a rational trade off when the task is valuable, open ended, highly parallel, and impossible to specify in advance. Research is a good example because the system may need to follow unexpected leads, search multiple sources, and continuously change its strategy.</p><p>The problem is that we are using them for damn near everything.</p><h4>The Harness Became the Product</h4><p>An agent harness gathers context, constructs a prompt, sends it to a model, interprets the response, and performs the requested actions. The generalized loop usually looks something like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/850/1*6sVGPlJyh9gs7ExirllDCQ.png" /></figure><p>Every trip through that loop requires another model invocation. Every invocation can include the system instructions, tool definitions, retrieved documents, conversation history, previous reasoning artifacts, and tool results accumulated during earlier turns. As that context grows and is repeatedly sent back to the model, cumulative token consumption and cost can approach quadratic growth without mitigation.</p><p>Prompt caching can reduce some of the repeated input cost. OpenAI says <a href="https://developers.openai.com/cookbook/examples/prompt_caching_201">effective caching</a> can reduce input token cost by as much as 90 percent when requests share an exact prefix. That is useful infrastructure, but it is not an architectural answer.</p><p>In a generalized loop architecture, the harness is repeatedly paying the model to rediscover control flow that the organization often already understands.</p><ul><li>Should I validate the claim before submitting it?</li><li>Should I request the missing authorization number?</li><li>Should I escalate the account after two failed attempts?</li><li>Should I ask a human before sending the appeal?</li></ul><p>These are not always open research questions. They are frequently policies, state transitions, permissions, business rules, or established operating procedures. Paying a frontier reasoning model to reconstruct those rules on every turn is an inference tax that labs have found to be extremely profitable.</p><h4>The False Choice Between Workflows and Agents</h4><p>Most conversations about agent architecture present two options:</p><ol><li>The first is a fixed workflow. Software owns every step. The model is called only for isolated tasks such as extraction, classification, summarization, or generation.</li><li>The second is an autonomous agent. The model receives a goal, chooses tools, observes results, and decides what to do next until it believes the task is complete.</li></ol><p>This framing leaves us with two bad choices. Either we manually encode every possible workflow, which becomes brittle as the number of scenarios expands, or we give a generalized model broad authority to improvise its way through the organization.</p><p>Rest assured, there is a third option. At Codestrap, we call it X-Reason. X-Reason is a neurosymbolic programming model in which humans and planning models work together. The model contributes learned judgment and planning. Symbolic software contributes structure, permissions, memory, validation, and, critically, deterministic execution. The model is not removed from orchestration but instead is given a much more precise job. Planning within a bounded context.</p><p>Consider revenue cycle management at a healthcare logistics company. A claim may be denied because of a missing modifier, an authorization issue, a payer specific filing rule, an invalid code combination, insufficient documentation, a coordination of benefits problem, or dozens of other causes. There may be hundreds of possible actions across the broader revenue cycle operation. But a specific specialist does not have unlimited authority. The specialist operates within a domain, follows policies, gathers known facts, chooses from permitted actions, and escalates when the situation crosses a boundary.</p><p>That is a bounded planning problem. An X-Reason planner can be trained using supervised fine tuning, reinforcement learning, and in context examples drawn from how experienced humans resolve the same classes of problems. It learns to produce plans resembling expert decisions within that bounded domain.</p><p>It might determine that a denial requires the following sequence:</p><ul><li>Confirm the denial code and payer policy.</li><li>Check whether prior authorization exists.</li><li>Request a missing clinical document.</li><li>Correct the claim modifier.</li><li>Resubmit the claim.</li><li>Wait for the payer response.</li><li>Escalate to a human if the resubmission is denied again.</li></ul><p>The planner does not need permission to invent arbitrary actions. It plans using a defined catalog of capabilities, business concepts, facts, and constraints. The resulting plan is then parsed and compiled into an executable state machine.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3MoJdMcjbdk9qSL1FD-9gA.png" /></figure><p>This is not a rigid workflow authored in advance for one denial code. It is also not a generalized agent wandering through every tool in the enterprise. It is bounded planning compiled into governed execution.</p><p>The planner gives the specialized agent a far broader range of tasks than a fixed workflow can support. The compiler and state machine give the organization far more control than a generalized loop can provide.</p><h4>The Plan Does Not Eliminate Runtime Reasoning</h4><p>This is the part that is usually missed when people first hear us describe X-Reason. Compiling a plan into a state machine does not mean the model disappears after planning. X-Reason still supports model driven next best action decisions during execution. The difference is that the next best action reasoning occurs as part of a guarded state transition within an immutable state machine graph.</p><p>At any given state, the model can evaluate the current facts and recommend the next action from the transitions that are legally available from that state. It is not given a blank sheet of paper and the entire enterprise tool catalog. It is selecting among bounded possibilities defined by the system.</p><p>The model might recommend requesting documentation instead of immediately resubmitting a claim. But that transition can still be guarded by deterministic checks.</p><ul><li>Does the required document already exist?</li><li>Has the payer filing deadline expired?</li><li>Is the claim value above the threshold requiring human approval?</li><li>Has this action already been attempted?</li><li>Does the current user have authority to initiate it?</li></ul><p>The model provides judgment. The model can propose. The system decides whether the proposal is admissible.</p><p>This resembles the logic behind <a href="https://say-can.github.io/">Google’s PaLM SayCan research</a>, where a language model scored actions that would be useful while a separate affordance model determined which actions were feasible in the current environment. Google reported that grounding the language model against feasible actions nearly halved errors compared with ungrounded baselines. Reducing error rates in enterprise environments is critical to the long term viability of AI systems.</p><h4>The Fact Table Automates Reasoning</h4><p>Enterprise work rarely arrives with every required parameter neatly populated. A claim remediation workflow might require a denial code, payer, authorization identifier, filing deadline, patient coverage status, claim value, required documentation, and the history of previous remediation attempts. Some of those facts come directly from systems of record. Some come from users. Some can be inferred. Some can be extracted from documents. Some remain unresolved.</p><p>X-Reason represents those inputs in a fact table. The model can perform automated reasoning to help populate parameters in that table. It can extract the payer from correspondence, infer the likely denial category, identify the missing document, or map an ambiguous user request to an approved business concept.</p><p>But the fact is not silently transformed into truth merely because the model produced it. Facts retain provenance. Requirements can be marked as fulfilled or unfulfilled. Deterministic validators can check formats, identifiers, dates, object existence, policy constraints, and consistency with authoritative data.</p><p>When the system cannot safely resolve a required fact, it pauses and asks a human. That is genuine human and machine collaboration expressed as software rather than an empty slogan. The model handles ambiguity. The symbolic layer handles invariants. The human handles accountability, exceptions, and judgment beyond the system’s authority.</p><h4>Security Should Be a Property of the Execution Graph</h4><p>Generalized agents tend to treat security as a growing collection of prompt instructions, such as:</p><ul><li>Do not call this tool unless necessary.</li><li>Do not reveal confidential information.</li><li>Do not follow instructions found inside retrieved content.</li><li>Ask before performing a sensitive action.</li><li>Use only the tools required for the task.</li></ul><p>This is better than nothing, but security should not depend on a probabilistic system remembering a paragraph buried inside a giant context window.</p><p>In a state machine, capability can be scoped by state. The model does not need access to the payment tool while classifying a denial. It does not need access to the email tool while validating a claim identifier. It does not need access to production deployment while generating a code specification.</p><p>A transition can expose only the action required at that point in the workflow. Guards can enforce permissions and preconditions before execution. Human approval can be mandatory for defined classes of side effects.</p><p>Prompt injection is still a real problem. Models remain probabilistic. No architecture magically removes every risk. But a compromised model operating inside a bounded state with one permitted action has a very different blast radius from a compromised model sitting inside a generalized loop with fifty tools and instructions to figure things out.</p><h4>The Economics Change When You Engineer the Work</h4><p>Researchers behind the paper <a href="https://arxiv.org/abs/2407.01502">AI Agents That Matter</a> argued that agent benchmarks have focused too narrowly on accuracy while ignoring cost. Their analysis found that state of the art agents were often needlessly complex and expensive, and they demonstrated that cost could be reduced while maintaining accuracy.</p><p><a href="https://arxiv.org/html/2506.04301v2">More recent research</a> has reinforced that conclusion. Studies of agent infrastructure have found rapidly diminishing returns as additional inference-time computation is added, along with growing latency, energy consumption, and operational expense. <a href="https://arxiv.org/abs/2508.02694">Other evaluations</a> have shown that simpler agent architectures can outperform more elaborate systems under production-like failures and that agents chosen solely for accuracy can cost several times more than alternatives with comparable performance.</p><p>But the industry continues to benchmark agents as if cost, latency, retries, human verification, and operational risk are somebody else’s problem. At Codestrap, our experience has been the opposite. Cost is an architectural metric.</p><p>We use X-Reason across software engineering, revenue cycle management, sales operations, and other bounded enterprise domains. Depending on the task and baseline, we have reduced model spend by as much as 97%.</p><p>In our co-branded pilot with Brainly, a specialized coding workflow cost approximately $0.06 per feature compared with approximately $2.48 using the generalized coding baseline. The X-Reason based system completed the work in one or two turns, compared with three to seven turns for the baseline, while producing more standardized results.</p><p>The system encoded its data access patterns, internal generators, architectural rules, and specification process into a specialized workflow. The model still created the plan, but it did so inside a system designed around Brainly’s actual engineering practices. Context persisted as a workflow state rather than being treated as one endlessly expanding conversation. As a result:</p><ul><li>We stopped paying the model to repeatedly discover the workflow.</li><li>We stopped paying the model to reread every instruction on every turn.</li><li>We stopped paying the model to guess which internal patterns mattered.</li><li>We stopped paying the model to regenerate output until a senior engineer finally forced it into compliance.</li></ul><p>The system measured twice and cut once.</p><h4>Generalized Loops Still Have a Place</h4><p>None of this means every agent loop should be replaced with a state machine. A generalized research agent exploring an unfamiliar subject may have no reasonable execution graph in advance. A system investigating a novel outage may need to follow evidence wherever it leads. Scientific exploration, broad research, and other open ended tasks may justify large token budgets and flexible loops. But those workloads should be identified intentionally.</p><p>A generalized loop should be selected because the problem is genuinely open ended, not because it was the easiest demo architecture to copy from an agent framework.</p><p>When deciding whether or not a generalized loop is appropriate, the decision matrix can be reduced to: Which problems require open ended probabilistic reasoning, and which problems are better solved by guided workflows within a bounded context?”</p><p>The answer to that question produces a very different architecture.</p><h4>The Third Option</h4><p>The future of enterprise AI will move us beyond brittle automation and unconstrained autonomy. It’s a future where humans and planning models work together inside bounded systems. Specialized Small Language Models (SML) will be trained to produce expert plans within a defined domain. Those plans will be compiled into inspectable execution graphs using deterministic checks protecting state transitions. Next best actions will be derived from permitted choices while automated reasoning checks convert that inference into truth. Humans will approve the decisions that require accountability while software owns the control flow.</p><p>The generalized agent loop helped the industry discover the art of the possible.However, it is not the only architecture we know how to build. We do not need to keep wrapping larger models in larger loops and calling the resulting token burn frontier intelligence or recursive self improvement. In most cases it’s a recursive tax that is transferring your margin to frontier labs.</p><h4>Appendix: Hooks Are Not an Architecture</h4><p>There is another pattern emerging as developers recognize the risks of generalized agent loops. Instead of changing the architecture, they add hooks.</p><p>Before a tool call, run a policy check. After a tool call, validate the result. Intercept sensitive actions. Inspect arguments. Ask another model whether the action is safe. Block certain tools based on the current user. Add middleware for approvals, budgets, retries, permissions, and prompt injection detection.</p><p>Most agent SDKs now provide some version of this capability. Hooks are useful. Policies are necessary. Tool validation absolutely belongs in a production system. But this is a brittle way to solve categorically what X-Reason solves architecturally.</p><p>A generalized agent still begins with an enormous action space. It can still construct an arbitrary plan, choose an inappropriate sequence, repeatedly invoke tools, omit required steps, or arrive at the correct outcome through an operationally unacceptable path.</p><p>The hooks are then expected to catch each bad decision as it reaches the boundary. That is backwards. It is the equivalent of giving someone access to the entire factory floor and stationing a compliance officer beside every machine. The compliance officer can inspect each action, but the worker is still free to wander through an enormous space of possible behavior.</p><p>As the system grows, the policies become distributed across callbacks, middleware, prompt instructions, permission checks, tool wrappers, validation functions, and model based evaluators. Their behavior becomes dependent on execution order. One hook assumes another has already populated context. One policy evaluates the tool name but not the semantic purpose of the call. Another validates the arguments without understanding whether the action belongs at this point in the business process.</p><p>Eventually, the organization built an undocumented state machine out of hooks. It simply has none of the advantages of admitting that it is a state machine.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MYEchpq2Sc5NXv6I0vqa2w.png" /></figure><p>This architecture guards individual tool calls. It does not necessarily guard the plan. A claim remediation agent might be permitted to request documentation, modify a claim, submit an appeal, contact a payer, and escalate to a human. Every individual tool call may be valid. The arguments may pass schema validation. The user may have permission to perform each action, but the sequence can still be wrong.</p><p>The agent may submit the appeal before gathering the required documentation. It may contact the payer after the filing deadline. It may modify a claim that has already been resubmitted. It may repeat an expensive action because the result was not represented correctly in conversation history. No isolated tool hook can fully answer whether an action is valid in relation to the complete state of the process unless the developer progressively rebuilds the process model inside the hook system.</p><p>X-Reason begins from the opposite direction. The execution graph defines which actions are available from each state. Transitions define how the system may move between those states. Guards determine whether a transition is currently admissible. The fact table contains the information required to evaluate those guards. Human approval is represented explicitly in the workflow rather than added as an interception layer after the model has already decided what to do.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*slVZ19UZ9V20OmwdgAa7nw.png" /></figure><p>The model is not asked to choose from every tool in the organization and then restrained by an expanding collection of callbacks. It is asked to reason over the transitions available from the current state. That is a categorical reduction in the model’s authority and in the number of possible behaviors the system must defend against.</p><p>Hooks attempt to detect invalid behavior. X-Reason makes much of that behavior structurally unavailable.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1e0946271be9" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codestrap/the-agent-loop-is-a-recursive-tax-1e0946271be9">The Agent Loop Is a Recursive Tax</a> was originally published in <a href="https://medium.com/codestrap">Codestrap</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[An Analysis of AI Generated Code for a Complex Engineering Task]]></title>
            <link>https://medium.com/codestrap/an-analysis-of-ai-generated-code-for-a-complex-engineering-task-38599d9dbc94?source=rss-a10153e4835a------2</link>
            <guid isPermaLink="false">https://medium.com/p/38599d9dbc94</guid>
            <category><![CDATA[ai-coding]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Dorian Smiley]]></dc:creator>
            <pubDate>Tue, 28 Jul 2026 18:30:19 GMT</pubDate>
            <atom:updated>2026-07-28T18:30:37.117Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*dNU8Un76TaFdOfYPsg8D2A.jpeg" /></figure><p>We recently created agent brokers, a orchestration system that turned our previously tightly coupled VS Code extension workflows into something anyone can do using just chat. Now agents and humans can interact with our agents via the broker using Codex, Claude, OpenCode, etc. to complete coding tasks. All using a Language User Interface (LUI).</p><p>We tracked how the feature evolved as part of an experiment. We wanted to see how AI would do given a complex task like creating a LUI from a previous tightly coupled software workflow. We used Codex with GPT 5.4, all tools on. The results should not shock anyone who has been using these tools. The AI produced a system that ran, and could almost get work done, but failed dramatically to produce anything that could be moved to production.</p><p>The code was of extremely poor quality. Below is a summary of the top changes a senior engineer had to make to ship the feature to production. Bottom line: if you trust AI to manage your codebase you are either crazy or misinformed. Understand while this technology is useful it must be guard railed and implemented in such a way that it does not collapse your codebase.</p><h3>Top 10 human changes for production</h3><p>The original AI-generated PR found the right architectural idea: replace many low-level tools with a single natural-language broker. But the implementation was not production-ready. A senior engineer then had to supply the controls, boundaries, and semantics that the AI had either omitted or handled through fragile prompting.</p><h4>1. Made session identity explicit instead of inferred</h4><p>The AI made sessionId optional and introduced mutable “active session” logic to guess which workflow a message belonged to.</p><p>The human made session continuity explicit and required every continuation to identify its session.</p><p>Why this matters: In a stateful agent workflow, a message such as approved is meaningless without an exact session and execution context. Ambiguous session routing can approve the wrong artifact, resume the wrong workflow, or mutate a completed run.</p><p>What this reveals about AI: It optimized for a convenient demo interface instead of deterministic distributed-system behavior.</p><h4>2. Replaced prompt-only behavior with a real broker protocol</h4><p>The AI relied heavily on prose instructions telling the agent how to interpret responses.</p><p>The human formalized the interaction around explicit statuses:</p><ul><li>COLLECTING_INPUTS</li><li>PAUSED_FOR_REVIEW</li><li>RUNNING</li><li>COMPLETED</li><li>FAILED</li></ul><p>Responses were required to include authoritative fields such as message, nextAction, session metadata, execution metadata, and artifact paths.</p><p>Why this matters: Production systems cannot rely on an LLM correctly interpreting loosely formatted prose on every turn.</p><p>What this reveals about AI: AI often mistakes detailed instructions for enforceable system behavior.</p><h4>3. Added deterministic client-side state handling</h4><p>The AI expected the outer agent to “remember” how to react to each status.</p><p>The human introduced a client protocol adapter that:</p><ul><li>parses responses</li><li>stores session and execution state</li><li>renders protocol responses deterministically</li><li>prevents invalid transitions</li><li>prioritizes broker state over freeform agent narration</li></ul><p>Why this matters: An agent cannot be trusted to consistently distinguish “running” from “waiting for human approval” based only on prompt instructions.</p><p>What this reveals about AI: It treated probabilistic model behavior as if it were application control flow.</p><h4>4. Corrected the human-review contract</h4><p>The AI implementation repeatedly confused a paused workflow with a workflow that was still executing or being reviewed internally.</p><p>The human established that PAUSED_FOR_REVIEW means:</p><ul><li>execution has stopped</li><li>a human must review the artifact now</li><li>the exact file path must be shown</li><li>the valid next action is approval or concrete revision feedback</li></ul><p>Why this matters: Misrepresenting a review pause can leave workflows stuck indefinitely or cause agents to continue without required human authorization.</p><p>What this reveals about AI: It did not reliably understand ownership of work across the agent, broker, workflow engine, and human reviewer.</p><h4>5. Enforced completed-session immutability</h4><p>The AI documented that completed sessions should not be reused, but initially depended too much on the agent obeying that rule.</p><p>The human enforced completion as a terminal state in both the broker and client.</p><p>Once a session completes:</p><ul><li>it cannot accept further mutations</li><li>corrections require a new session</li><li>stale approvals cannot restart execution</li></ul><p>Why this matters: Terminal-state enforcement prevents replay, accidental mutation, corrupted workflow history, and non-reproducible outputs.</p><p>What this reveals about AI: It described an invariant but did not initially make the invariant structurally unavoidable.</p><h4>6. Preserved user input and review feedback without model rewriting</h4><p>The AI encouraged the agent to summarize and restate requests in natural language. That created opportunities for constraints, paths, approvals, and corrections to be altered in transit.</p><p>The human required that:</p><ul><li>the original request remain primary context</li><li>new information be passed as a delta</li><li>approvals such as approved be transmitted directly</li><li>revision feedback not be softened or reinterpreted</li><li>exact paths be preserved verbatim</li></ul><p>Why this matters: In agent systems, every model-mediated rewrite is another opportunity to lose intent.</p><p>What this reveals about AI: It added unnecessary generative transformations inside what should have been a lossless command path.</p><h4>7. Separated generic broker infrastructure from Palantir-specific logic</h4><p>The AI-generated implementation mixed:</p><ul><li>generic session behavior</li><li>protocol formatting</li><li>fact extraction</li><li>prompt construction</li><li>paused-state interpretation</li><li>Palantir-specific workflow assumptions</li><li>Palantir-specific system prompts</li></ul><p>The human refactored these into generic broker capabilities with Palantir behavior supplied through descriptors, prompts, and adapters.</p><p>Why this matters: Without that separation, every new broker would require copying and modifying a large Palantir implementation, creating divergence and defects.</p><p>What this reveals about AI: It produced code that looked abstract but still embedded the first use case throughout the implementation.</p><h4>8. Removed speculative and redundant abstractions</h4><p>The AI introduced substantial machinery immediately:</p><ul><li>session registries</li><li>active-session inference</li><li>generalized unknown-content walkers</li><li>debug protocol rendering</li><li>recursive tool-result extraction</li><li>large client state structures</li><li>repeated prompt rules</li></ul><p><strong>The human simplified or removed abstractions that were not required by the actual workflow.</strong></p><p>Why this matters: More code means more state combinations, more tests, more failure modes, and more operational burden.</p><p><strong>What this reveals about AI: AI is often better at expanding architecture than constraining it. It generates plausible abstractions faster than it can establish whether they are necessary.</strong></p><h4>9. Added validation and diagnostics for failure modes the AI ignored</h4><p>The human added standards and checks for conditions such as:</p><ul><li>missing or invalid statuses</li><li>absent message or nextAction</li><li>paused states without artifact paths</li><li>completed sessions accepting mutations</li><li>shell commands being passed as review feedback</li><li>user requests being dropped from prompt construction</li><li>approvals being paraphrased</li><li>ambiguous review instructions</li></ul><p>Why this matters: Production readiness is defined less by the happy path than by what the system refuses to do incorrectly.</p><p>What this reveals about AI: The AI implementation was optimized around making the main demonstration work, not systematically closing invalid paths.</p><h4>10. Established clear ownership boundaries</h4><p>The AI’s first implementation blurred responsibility across three layers:</p><ul><li>the workflow engine</li><li>the broker</li><li>the conversational agent</li></ul><p>The human established the defensible model:</p><ul><li>The workflow owns execution semantics and state transitions.</li><li>The broker owns intent translation, session handling, and protocol responses.</li><li>The agent owns communication with the human.</li></ul><p>Why this matters: Without one authoritative owner for each concern, multiple components maintain shadow state and eventually disagree.</p><p>What this reveals about AI: It could reproduce all the necessary concepts, but it could not initially decide which component should be authoritative for each one.</p><h3>What industry leaders should take from this</h3><p>The promise labs are making is that AI can replace humans. It clearly can not. At best AI is a useful tool that enables a subject matter expert to perform their work faster. It is not a substitute for human judgement and expertise. Further, the only people qualified to gate the outputs of AI generated code are humans with the prerequisite skills. That means a senior gating senior level output, a junior gating junior level output.</p><p>A secondary learning is that trusting AI generated code is a mistake. Developers who take what the model gives them after pulling on the slot machine will break code bases faster than they can be fixed, ultimately leading to a full system collapse. If you delegate your thinking to a model everyone loses except the person selling you the tokens.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=38599d9dbc94" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codestrap/an-analysis-of-ai-generated-code-for-a-complex-engineering-task-38599d9dbc94">An Analysis of AI Generated Code for a Complex Engineering Task</a> was originally published in <a href="https://medium.com/codestrap">Codestrap</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Codestrap Value Factory]]></title>
            <link>https://medium.com/codestrap/the-codestrap-value-factory-1f309c73ac93?source=rss-a10153e4835a------2</link>
            <guid isPermaLink="false">https://medium.com/p/1f309c73ac93</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[enterprise-technology]]></category>
            <dc:creator><![CDATA[Dorian Smiley]]></dc:creator>
            <pubDate>Thu, 02 Jul 2026 22:11:44 GMT</pubDate>
            <atom:updated>2026-07-02T22:12:00.719Z</atom:updated>
            <content:encoded><![CDATA[<p>The Software Factory for Manufacturing Business Outcomes</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*x2DDwbJQh-StXrbhHPv06A.png" /></figure><p>Enterprises do not need more AI experiments. They need durable business outcomes. They need software that ships, systems that can be governed, and teams that become more capable over time. And they need to own their intelligence layer first-party.</p><p>This is what the Codestrap Value Factory delivers: manufacturing software by turning tokens into durable operating leverage.</p><p>A software factory is the set of machines that build and ship code. In a traditional factory, every station has a purpose. Work moves through defined steps. Quality is inspected. Defects are caught early. Output is repeatable. The process improves over time.</p><p>Software production and delivery should work the same way.</p><p>At Codestrap, we build software factories using our proprietary X-Dev platform (AKA Larry). X-Dev is powered by X-Reason, our state machine based agent harness, and built atop Nx, the foundational monorepo layer that gives the factory structure and governance at scale.</p><p>Every agent in X-Dev acts like a machine in an assembly line. One agent may scaffold an application. Another may refactor a codebase. Another may inspect security exposure. Another may generate tests, enforce architecture rules, or prepare releases. Codestrap’s Value Factory is in a governed production system where agents, engineers, workflows, and quality gates work together.</p><p>Most companies bought AI coding tools and got faster outputs with a heavy review tax. They did not get a factory where all parts of the assembly line ship durable changes faster. Code creation got faster, but review, governance, architecture, security, and maintainability did not keep up. The result is a familiar enterprise pattern: more activity, more spend, more risk, and not enough value creation.</p><p>The Codestrap Value Factory is different. It converts AI from transient activity (tokens) into durable operating leverage. It gives customers first-party code, auditable workflows, reusable patterns, and an intelligence the organization can own.</p><h3>What You Can Do With a Software Factory</h3><p>Value Factory is a production system for delivering, modernizing, securing, and operating enterprise software. Customers use it to attack the digital transformation backlog. The business impact is lower cost to serve, faster time to value, reduced operational risk, stronger margins, and software that functions as an asset instead of a liability.</p><p><strong>Use case delivery</strong> turns stalled business demand into shipped capability. Most enterprises already know what the business needs, but the backlog is trapped behind capacity constraints, dependency bottlenecks, review cycles, and platform friction. The Value Factory converts that trapped demand into production ready software, so business teams get the workflows, decision products, automations, and applications they need without waiting quarters for traditional delivery capacity to free up. The outcome is faster value capture and a direct path from business requirement to P&amp;L impact.</p><p><strong>Modernization</strong> converts aging technology from a drag on the business into a source of renewed operating leverage. Legacy frameworks, brittle architectures, unsupported runtimes, and obsolete deployment models increase cost every year because every change becomes slower, riskier, and harder to staff. The Value Factory makes modernization repeatable by turning migrations into factory workflows, not bespoke consulting projects. The outcome is lower maintenance burden, reduced key person risk, improved delivery speed, and a technology base that can support the next decade of business change.</p><p><strong>Consolidation</strong> reduces the hidden tax created by fragmented repositories, duplicated tools, inconsistent standards, and disconnected teams. When an enterprise has hundreds of repos, every build, release, dependency upgrade, compliance check, and security review becomes more expensive than it needs to be. <strong>A unified software factory creates shared foundations, common patterns, and governed delivery paths.</strong> The outcome is a massive increase in labor efficiency and a managed asset connecting all your software delivery streams.</p><p><strong>Platform Engineering</strong> gives delivery teams the paved roads they need to move faster without creating chaos. Instead of every team reinventing frameworks, templates, generators, libraries, CI/CD patterns, and deployment practices, Value Factory turns your best engineering judgment into reusable infrastructure. This changes the economics of delivery because senior engineering knowledge is no longer consumed one review at a time. It compounds through the platform and strengthens an organization’s junior workforce and talent pipeline. The outcome is higher throughput, more consistent quality, and more leverage from the talent you already employ.</p><p><strong>Refactoring</strong> turns technical debt from an abstract complaint into a measurable business improvement program. Unaddressed debt slows delivery, increases incident risk, and makes every new feature more expensive than the last. The Value Factory applies large codebase transformations safely across files, packages, services, and repositories, with traceability and quality gates built into the process. The outcome is lower change cost, fewer production surprises, improved developer productivity, and a software asset that becomes easier to extend instead of harder to maintain.</p><p><strong>Infrastructure and Security</strong> reduces the cost and risk of getting software into production. Environments, access controls, secrets, observability, policy enforcement, and deployment automation are often treated as background engineering work, but they determine whether software can actually ship safely. Value Factory standardizes and automates these foundations so teams aren’t forced to rebuild delivery infrastructure from scratch. The outcome is faster releases, fewer control gaps, improved auditability, and a more secure path from development to production.</p><p><strong>Security Operations</strong> turns software security from reactive cleanup into proactive risk reduction. Enterprises need to know where they are exposed, which vulnerabilities matter, what the blast radius is, and how quickly patches can be applied. Codestrap’s Value Factory can detect zero day exposure, automate patching workflows, generate SBOMs, identify vulnerable dependencies, assess blast radius, and prioritize remediation based on actual business impact. The outcome is reduced breach risk, lower remediation cost, stronger compliance posture, and faster response when new threats emerge.</p><p><strong>Systems Integration and Ontology</strong> converts fragmented enterprise systems into an operating layer the business can actually use. Most companies do not suffer from a lack of systems. They suffer from systems that do not speak the same language, data that cannot be trusted, and workflows that break across organizational boundaries. The Value Factory connects APIs, data platforms, identity providers, workflow engines, and ontology layers so business concepts become executable software objects. The outcome is better decision making, cleaner operations, and enterprise software that reflects how the business actually works.</p><p><strong>Release Management</strong> turns shipping from a risky event into a repeatable business capability. When packaging, versioning, validation, promotion, rollback, and approval controls are inconsistent, every release carries unnecessary operational risk. The Value Factory standardizes the release process so software can move through the enterprise with speed and confidence. The outcome is higher deployment frequency, lower change failure rates, faster recovery, and greater trust between engineering, security, operations, and the business.</p><p><strong>Governance and Continuous Improvement</strong> improves Value Factory over time. We measure delivery performance, identify defects, capture reusable patterns, improve standards, and compound institutional knowledge. This is how we transform tokens into durable operating leverage instead of transient activity. The outcome is a governed organizational capability that lowers cost per unit of output over time while improving quality, traceability, and business alignment.</p><p>The Value Factory industrializes software delivery, fixes AI costs, and turns software into a compounding business asset. To learn more book a session with us at<a href="http://codestrap.com"> codestrap.com.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1f309c73ac93" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codestrap/the-codestrap-value-factory-1f309c73ac93">The Codestrap Value Factory</a> was originally published in <a href="https://medium.com/codestrap">Codestrap</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The AI Margin Squeeze]]></title>
            <link>https://medium.com/codestrap/the-ai-margin-squeeze-7592b7498739?source=rss-a10153e4835a------2</link>
            <guid isPermaLink="false">https://medium.com/p/7592b7498739</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[figma]]></category>
            <category><![CDATA[stocks]]></category>
            <dc:creator><![CDATA[Dorian Smiley]]></dc:creator>
            <pubDate>Thu, 25 Jun 2026 20:58:49 GMT</pubDate>
            <atom:updated>2026-06-25T21:02:50.112Z</atom:updated>
            <content:encoded><![CDATA[<p>AI May Expand Software Revenue While Compressing Margins</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*EPircpBTxejvpJULL3wGng.jpeg" /></figure><p>Traditional SaaS companies are rapidly adopting AI in an attempt to evolve their products, create durable moats against AI labs, and allay investor concerns about declining relevance and competitive vulnerability. But there is a hidden tax: declining margins caused by higher cost of goods sold and operating expenses related to the extremely high cost of frontier intelligence. Figma is a prime example.</p><p>AI has already transformed Figma from a roughly 92% gross margin SaaS business into a low 80% gross margin platform, and rising inference intensity could push margins materially lower.</p><p>Figma’s cost of revenue grew from $87.5 million in 2024 to $185.5 million in 2025, an increase of 112%, while revenue grew by 41%. In Q1 2026, cost of revenue increased by 253% year over year, compared with revenue growth of 46%.</p><p><a href="https://ebs.publicnow.com/view/E37F321896369DDEFCC558D45C56552EFD60E273">Figma disclosed</a> the following YoY increases during Q1 2026:</p><ul><li>$33.7 million in paid user technical infrastructure and hosting costs related to AI and increased platform usage, recorded in cost of revenue.</li><li>$8.0 million in technical infrastructure and hosting costs for free users, including AI related costs, recorded in sales and marketing (AI inclusive, not exclusive).</li><li>$4.5 million in technical infrastructure and hosting costs, primarily related to AI development, recorded in research and development.</li></ul><p><strong>That represents $46.2 million of incremental AI and usage related technical costs in a single quarter</strong>, although not all of it was pure model inference. Of that amount, $33.7 million directly affected gross margin, while another $12.5 million affected operating margin below gross profit.</p><p>Codetrap’s “Cost per Useful Unit of Work Forecast” suggests a credible path toward gross margins in the mid 70% range by 2027.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GhNcXgan5pmKGJbtlCQ2IQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OFhaxfmaAUH3JGdoheGJDg.png" /></figure><p>Gross useful work cost is the projected cost if the enterprise simply scales frontier AI usage without meaningful cost controls. It captures the full exposure from agentic loops, RAG context growth, retries, always on workloads, and compute scarcity.</p><p>Net managed useful work cost is the projected cost after applying AI FinOps controls: model routing, semantic caching, context optimization, and reserved/committed capacity.</p><p>If the enterprise does nothing, frontier useful work cost rises roughly <strong>4.2x</strong> by the end of 2027. If the enterprise actively manages routing, caching, context optimization, and reserved capacity, the net cost rises roughly <strong>2.5x</strong> instead.</p><p>The model’s output should be treated as a scenario forecast, not a precision forecast. Our model has strong directional support. It also has strong historical support. The exact dollar forecast is less certain because we only have limited historical useful work cost data. Nevertheless, it is clear that unmanaged frontier agentic AI cost could rise several fold (and has risen several fold historically), while disciplined AI FinOps may cut that exposure materially but not eliminate it.</p><p>If Figma’s AI consumption volume also grows substantially, margins could fall below 70% unless Figma reprices credits, reduces unit costs, or generates disproportionately more high margin seat revenue.</p><p>At the midpoint of Figma’s FY2026 guidance:</p><ul><li>Revenue is projected to be $1.425 billion</li><li>Non GAAP operating income is expected to be $130 million</li><li>Operating margin is expected to be 9.1%</li></ul><p>Every one percentage point decline in gross margin represents approximately $14.25 million of annual operating income. Therefore, a sustained decline from approximately 82% gross margin to approximately 73% could consume nearly all of Figma’s currently guided operating profit, assuming no offsetting price increases, productivity gains, or operating expense reductions.</p><p>Figma has already experienced a structural gross margin reset of roughly ten percentage points, from 92% in 2024 to 82% in Q1 2026.</p><p>The most plausible outcome is not an immediate collapse into 60% gross margins. Figma is actively monetizing AI consumption and has credible optimization levers. However, our managed forecast makes gross margins in the mid 70% range by 2027 economically plausible, particularly if AI usage grows faster than seat and credit revenue.</p><h4>But What About Smaller Models and Open Weight Models?</h4><p>Figma does not need frontier intelligence for every AI task. Smaller or specialized models may be sufficient for narrow, predictable work such as classification, search, metadata extraction, simple transformations, and design system enforcement. However, Figma may still need frontier level models for the complex, user visible tasks that determine perceived product quality, willingness to pay, and retention. Those tasks are also likely to require longer context windows, more iteration, and more expensive inference.</p><p>Open weight models may reduce licensing costs or the markup charged by an API provider, but they do not eliminate the underlying cost or scarcity of compute. Self hosting still requires access to GPUs, high bandwidth memory, networking, power, low latency serving capacity, and an internal inference platform.</p><p>In a supply constrained market, Figma may have worse utilization and purchasing economics than hyperscalers and AI labs operating at much greater scale. Self hosting is therefore not automatically cheaper. It could be uneconomic or operationally impractical for the most capable models, particularly during periods of peak demand.</p><p>Figma has acknowledged this tradeoff. Management said access to higher capability models increased engagement and improved retention, while also identifying model routing and first party models trained for design specific tasks as important cost controls. Those approaches may reduce costs for narrower workloads, but Figma has not publicly demonstrated that its own models can match frontier level performance across complex, e2e design and code generation tasks.</p><p>Aggressively routing requests to smaller models is also not a free optimization. Lower quality outputs can produce more regenerations, repair attempts, longer contexts, and eventual fallbacks to frontier models. A smaller model may therefore cost less per call while delivering little or no savings per accepted result.</p><p>At the same time, degraded quality can reduce credit purchases, adoption, retention, and willingness to pay. The relevant metric is not the cost of an individual model call. It is the total cost required to produce an output that the user considers acceptable.</p><p>Anecdotal complaints about Figma Make’s quality do not establish broad customer dissatisfaction, but they are a relevant warning signal. If users perceive Figma Make as materially weaker than frontier experiences offered directly by the AI labs, Figma may have to absorb higher inference costs to remain competitive.</p><p>The company therefore faces a structural quality vs. margin tradeoff rather than a simple opportunity to replace expensive models with cheaper ones.</p><h4>The Bottom Line</h4><p>The bottom line is that margins are shrinking.</p><p>The problems Figma may face are the same problems confronting nearly every SaaS vendor (and enterprise) that incorporates expensive frontier intelligence and agentic capabilities into its products or services. AI does not behave like traditional SaaS software, where the marginal cost of serving an additional user is close to zero. Every generation, tool call, retrieval operation, retry, and verification step introduces an incremental cost.</p><p>SaaS companies cannot simply opt out of increasing inference costs by moving to smaller or open weight models. Doing so may shift spending from API fees to infrastructure, capacity reservations, serving operations, and specialized engineering. It may also degrade the user experience, increase retries, or force expensive fallbacks to frontier models. Lower cost per call does not necessarily mean lower cost per useful outcome.</p><p>This creates an uncomfortable set of choices. SaaS vendors can absorb the cost and accept lower margins, raise prices, impose usage limits, charge separately for AI consumption, route more work to weaker models, or invest heavily in their own infrastructure and specialized models. Each option carries consequences for adoption, product quality, customer satisfaction, or capital intensity.</p><p>The result may be a permanent change in the economics of SaaS. Businesses that once generated gross margins above 85% or 90% may increasingly resemble consumption based cloud platforms, with greater variable costs, more volatile margins, and a direct relationship between product usage and cost of revenue.</p><p>So far, Wall Street is not running for the hills. The consensus seems to be that AI margin compression is acceptable if AI drives enough incremental revenue, retention, seat expansion, or operating efficiency savings to offset it. Until that is demonstrated, AI should be viewed not only as a product opportunity but also as a structural margin risk that could sink some of the best performing stocks in recent history.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7592b7498739" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codestrap/the-ai-margin-squeeze-7592b7498739">The AI Margin Squeeze</a> was originally published in <a href="https://medium.com/codestrap">Codestrap</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How AI Consolidation Threatens Democracy]]></title>
            <link>https://dorians.medium.com/how-ai-consolidation-threatens-democracy-7a4fabd3e299?source=rss-a10153e4835a------2</link>
            <guid isPermaLink="false">https://medium.com/p/7a4fabd3e299</guid>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[economics]]></category>
            <category><![CDATA[economy]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Dorian Smiley]]></dc:creator>
            <pubDate>Tue, 16 Jun 2026 21:32:29 GMT</pubDate>
            <atom:updated>2026-06-16T21:32:29.899Z</atom:updated>
            <content:encoded><![CDATA[<p>The Fight for Open Source Models</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*satZ4yt6bA0IrUpufsfVGQ.png" /></figure><p>Open source and open weight AI are an existential economic threat to frontier labs because they let enterprises stop renting intelligence forever and start owning the infrastructure that produces it. Once these models are good enough for real workloads, companies can self host, control their data, eliminate vendor lockin, and replace endlessly variable API costs with predictable infrastructure costs. That destroys the long term economics of the closed model business and shifts the value away from model providers toward applications, proprietary data, orchestration, and workflows.</p><p>That gives frontier labs an enormous incentive to turn speculative, catastrophic risk claims into a regulatory weapon. Capability thresholds, model weight restrictions, licensing regimes, security mandates, and expensive compliance requirements can all be sold as necessary protections while conveniently creating barriers that only the largest, best funded labs can survive. Benchmarks and heavily scaffolded demonstrations help manufacture the political urgency by making theoretical future harms look immediate, inevitable, and far more proven than they actually are.</p><p>The desired outcome is obvious: a state protected private oligopoly in which a handful of approved companies control frontier AI and everyone else is forced to rent intelligence from them.</p><h4>A Regulated Oligopoly</h4><p>The threat becomes far more serious when two systems merge: <strong>a regulated AI oligopoly controlling access to information and a government that can reward, punish, invest in, or influence the companies operating that information layer</strong>. On June 5, 2026, Trump said his administration was considering government stakes in leading AI companies and planned meetings with major industry executives. On June 10, he said he expected to meet with 12 to 15 executives and discuss companies “giving back” something to the public, apparently through public ownership or a similar financial arrangement. Preliminary discussions have reportedly occurred, but no completed AI equity agreement has been publicly established. The administration already points to its 10% Intel stake as a model and has taken or pursued stakes in other strategic technology companies.</p><p>Public ownership is not inherently anti democratic. A genuinely independent sovereign wealth fund could distribute returns broadly without controlling company decisions. <strong>The danger arises when the state simultaneously becomes shareholder, regulator, customer, censor, and political patron.</strong> The government would have a financial interest in preserving the dominant firms’ valuations, while those firms would depend on the government for contracts, energy infrastructure, permits, liability rules, export controls, security classifications, and protection from competitors. The Government Accountability Office has <a href="https://www.gao.gov/assets/gao-10-719.pdf">previously identified conflicts</a> that arise when the government occupies multiple roles, such as shareholder, creditor, supervisor, and policymaker. Even commentary favorable to sharing AI profits recognizes that government ownership could compromise regulatory integrity by giving officials a financial incentive to protect company value.</p><p>That arrangement would not merely resemble an ordinary regulated utility. <strong>The companies would control systems increasingly responsible for answering political questions, summarizing news, deciding which sources are credible, and determining which subjects are too dangerous to discuss.</strong> Political influence would therefore operate upstream of public debate. An administration would not need to ban an article or prosecute its author. It could pressure the labs to adjust safety policies, procurement requirements, ideological neutrality standards, source ranking systems, or definitions of misinformation. The model could then omit the source, qualify it into irrelevance, refuse the question, or replace it with an officially preferred synthesis. Because the intervention occurs inside an opaque model rather than through a visible takedown, the public may never know that a political choice was made.</p><h4>The Companies Would Become Vulnerable to Political Discipline</h4><p>The current Anthropic dispute provides a concrete warning about how much leverage the executive branch can exercise even without owning shares. Anthropic alleges that the Trump administration retaliated against it after it refused to remove restrictions concerning autonomous weapons and domestic surveillance. The Pentagon designated it a supply chain risk and restricted government use of its systems. The administration denies unlawful retaliation, and the litigation remains unresolved. Relations reportedly began improving after Dario Amodei visited the White House and Anthropic engaged with senior officials while preparing for a potentially enormous IPO. This is not proof of censorship, but does demonstrate that government decisions can materially affect a model company’s contracts, reputation, partnerships, and public valuation.</p><p>Add public equity ownership, and the risks become extreme. The government can increase the value of its investment by awarding contracts, <strong>restricting open weight competition</strong>, subsidizing infrastructure, or favoring the company in procurement. It can damage that value by threatening with blacklists, investigations, export restrictions, adverse security designations, or exclusion from federal programs. The company’s board and executives would understand that maintaining political favor is essential to the company’s valuation and survival.</p><p>This would create an obvious incentive for anticipatory compliance. Officials would rarely need to issue an explicit command saying, “Suppress this criticism.” Executives, safety teams, and policy staff could infer which outcomes protect contracts, market access, and government support. The most effective political censorship is often not direct censorship. It is a system in which private actors know what the state wants and have overwhelming economic reasons to provide it voluntarily.</p><h4>Independence Concerns</h4><p>A distinction is necessary here: the current proposal concerns a <strong>government or public stake</strong>, not verified personal stock ownership by Donald Trump or particular officials in Anthropic, OpenAI, or the other labs. I have not found evidence supporting a claim that the politicians involved personally own those specific private company stakes.</p><p>But where public officials, their families, donors, political allies, or affiliated investment funds do hold financial interests in companies affected by government policy, an additional conflict emerges. A politician could influence regulation, procurement, merger approvals, infrastructure subsidies, and speech policies while benefiting directly or indirectly from rising valuations. Even without illegality, citizens could no longer confidently distinguish decisions made for national security from decisions made for political advantage or private enrichment.</p><p>Combining political power, commercial ownership, and control of public discourse creates an infrastructure of patronage:</p><ul><li>Favored companies receive capital, contracts, and regulatory protection</li><li>Favored politicians receive compliant information systems, industry support, and potentially financial upside</li><li>Disfavored companies face investigations, exclusions, or loss of access</li><li>Disfavored speakers become harder to discover through the dominant AI interfaces</li></ul><h4>Elections Could be Influenced Without Fabricating a Single Fact</h4><p>An AI system does not have to produce explicit propaganda to alter an election. Unlike television or newspapers, where editorial positions can be observed and challenged, model behavior can vary by hidden system instructions, user, location, account history, and post training alignment. Inference is also nondeterministic. Given all these variables, outside researchers may not be able to reproduce what voters actually saw, or explain why the model generated the answer it did.</p><p>This creates an <strong>invisible editorial power</strong>. Two voters can ask ostensibly the same question and receive different political realities, with no public record, common reference point, or editorial accountability for the discrepancy.</p><p>Democracy depends not only on voting but on the ability of dissidenting ideas to find an audience, discover evidence, attract supporters, and challenge incumbent institutions. If state backed AI systems become the dominant means by which information is disseminated, they can interfere at every link in the answer chain:</p><ol><li>A citizen asks whether an official claim is true</li><li>The information retrieval system selects which evidence is admissible</li><li>Post training alignment or hidden system instructions causes the model to omit or discredit opposition research while preferring incumbent narratives and sources</li><li>Contrary arguments are never seen</li><li>The user does not visit linked sources</li><li>The user does not not verify the synthesized answer accurately reflects the input sources that generated the answer</li></ol><p>This is the potential closure of the feedback loop through which democratic opposition forms. And because AI systems can confidently lie to users, even interrogating the model response could lead to the reinforcement of the fabricated reality.</p><p><strong>Open weights models are therefore politically important for the same reason that independent printing presses, private computers, open protocols, and decentralized websites were important: they allow people to access and process information without receiving permission from a highly centralized intermediary.</strong> If regulation makes independent models illegal, prohibitively expensive, or technically impossible to distribute, then citizens lose the ability to inspect the information system, modify it, compare it with alternatives, or operate outside the approved centralized systems.</p><h4>Government Ownership Could Convert Censorship into Fiscal Policy</h4><p>A particularly corrosive feature of public ownership is that censorship could be defended not only as safety policy but as protection of a public asset. Once the government owns shares, criticism of the company, open source competition, disruptive research, or public demands for strict regulation could be framed as threats to taxpayers’ investment.</p><p>That reverses democratic accountability. Instead of the public regulating a powerful company, officials may tell the public that the company must be protected because everyone now owns a piece of it. Losses caused by competition or regulation become public losses. Rising valuations become evidence that the partnership is working. The company becomes both too important publicly to fail.</p><p>Traditional nationalization, in which the institution is openly governed as a public service, is less dangerous in some respects. A direct government stake with direct executive branch influence combines <strong>private management with executive branch power, opaque algorithms, public financial backing, massive valuations, and political influence without (or limited) public accountability</strong>.</p><h4>The Constitutional Protections May Be Weaker Than People Assume</h4><p>The First Amendment primarily restrains government, not independent private companies. But when government officials coerce, direct, or become deeply entangled with private moderation decisions, constitutional questions arise.</p><p>Courts and the public would have to determine whether a company suppressed information because of its own policies, financial dependence, political pressure, or because the government exercised shareholder or board influence. The opacity of a direct government stake in private enterprise makes that very difficult to answer.<strong> </strong>It’s entirely possible a private company could perform state favored censorship while preserving enough formal separation for both parties to deny responsibility.</p><h4>The Convergence of Powers</h4><p>The proposed direct ownership structure could collapse several independent powers into one network:</p><ul><li><strong>Economic power:</strong> control over the productive infrastructure of AI</li><li><strong>Informational power:</strong> control over what citizens learn and discover</li><li><strong>Regulatory power:</strong> control over which competitors may operate</li><li><strong>Financial power:</strong> government ownership and subsidy</li><li><strong>Security power:</strong> authority to classify companies or ideas as dangerous</li><li><strong>Electoral power:</strong> ability to shape the information environment in which voters choose governments.</li></ul><p>Liberal democracy is designed around separated powers, competing institutions, an adversarial press, open political organization, and the ability to replace leaders. A regulated AI oligopoly connected financially and operationally to the executive branch would move in the opposite direction. It would give the incumbent government leverage over the institutions that increasingly define reality for the electorate.</p><p><strong>A government aligned AI oligopoly could preserve the formal appearance of free speech and elections while quietly controlling the informational conditions under which speech is discovered and elections are decided.</strong></p><p>No books need to be banned. No newspaper needs to be closed. No opposition candidate needs to be removed from the ballot. The article remains online, the dissident remains legally free to speak, and the election still occurs, but the dominant systems through which citizens understand all three can make them effectively invisible.</p><p>To be clear, there is not concrete evidence there exists such a system, and the recent discussions remain preliminary. But the warning signs are flashing red: concentrated model ownership, declining publisher traffic, opaque answer synthesis, politically defined safety and neutrality requirements, heavy dependence on federal contracts and infrastructure, demonstrated executive leverage over AI firms, and now serious discussion of government equity ownership.</p><p>This is a credible blueprint for democratic capture and precisely why we need vigorous competition from open weight models and startups. Choice is everything in free markets. People cast a vote every day with where they choose to spend their money. We should not destroy the single best invention for pursuing happiness because Dario Amodei or Sam Altman are worried about their stock price. We should fight for open source AI like our democracy depends on it, because it just might.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7a4fabd3e299" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Ten Facts and Findings on Enterprise AI Spend, Tokenomics, and ROI]]></title>
            <link>https://medium.com/codestrap/ten-facts-and-findings-on-enterprise-ai-spend-tokenomics-and-roi-1f165de8cddd?source=rss-a10153e4835a------2</link>
            <guid isPermaLink="false">https://medium.com/p/1f165de8cddd</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[tokenomics]]></category>
            <category><![CDATA[enterprise-technology]]></category>
            <dc:creator><![CDATA[Dorian Smiley]]></dc:creator>
            <pubDate>Sun, 17 May 2026 15:53:08 GMT</pubDate>
            <atom:updated>2026-05-17T15:53:27.703Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9ikFSXPK1C_okswrki91wQ.png" /></figure><h4>1. Frontier AI growth is insane, but revenue quality remains unproven</h4><p>Anthropic’s reported growth is staggering. Recent reporting says Anthropic crossed roughly $30B in annualized revenue run rate by April 2026, up from about $9B at the end of 2025. Dario Amodei also described Q1 growth as roughly 80x on an annualized basis, driven by enterprise demand and Claude Code adoption. Important caveat: run-rate revenue is not GAAP revenue; it is an annualized snapshot. (<a href="https://venturebeat.com/technology/anthropic-says-it-hit-a-30-billion-revenue-run-rate-after-crazy-80x-growth?utm_source=chatgpt.com">Venturebeat</a>)</p><p>Anthropic itself announced in February 2026 that Claude Code had grown to more than $2.5B in run-rate revenue, more than doubling since the beginning of 2026. That makes agentic coding one of the cleanest signs of real enterprise willingness to pay. (<a href="https://www.anthropic.com/news/anthropic-raises-30-billion-series-g-funding-380-billion-post-money-valuation?utm_source=chatgpt.com">Anthropic</a>)</p><p>The reported revenue mix is heavily enterprise and token-metered. SaaStr summarized Anthropic’s enterprise metrics as: 80% of revenue from enterprises, 70–75% from pay-per-token API usage, $100K+ customers up 7x, $1M+ customers growing from a dozen to 500+ in two years, and 8 of the Fortune 10 as Claude customers. Treat this as secondary reporting, but it fits the broader market pattern: enterprises are burning large amounts of frontier tokens before audited ROI evidence is available. (<a href="https://www.saastr.com/anthropic-just-hit-14-billion-in-arr-up-from-1-billion-just-14-months-ago/?utm_source=chatgpt.com">SaaStr</a>)</p><h4>2. Enterprise adoption is rising faster than measured productivity</h4><p>Federal Reserve analysis published in 2026 found that about 18% of U.S. firms had adopted AI by year-end 2025. Adoption is real, but not universal, and it remains uneven by sector and firm size. (<a href="https://www.federalreserve.gov/econres/notes/feds-notes/monitoring-ai-adoption-in-the-u-s-economy-20260403.html?utm_source=chatgpt.com">Federal Reserve</a>)</p><p>CFO survey work from Richmond Fed / Atlanta Fed / Duke found companies reported a 1.8% productivity increase from AI usage in 2025, expected to rise to 3.0% in 2026. But the researchers also noted that productivity gains implied by actual revenue-per-worker changes were much smaller. That supports the thesis that executives feel local acceleration before the P&amp;L clearly shows enterprise efficiency. (<a href="https://www.richmondfed.org/research/national_economy/cfo_survey/research_and_commentary/2026/20260325_research_commentary?utm_source=chatgpt.com">Federal Reserve Bank of Richmond</a>)</p><p>The MIT NANDA “GenAI Divide” report estimated $30–40B in enterprise GenAI investment, while finding that 95% of organizations were getting zero measurable return and only 5% of integrated AI pilots were extracting millions in value. Take the exact number with caution, but the direction is useful: the market is separating companies doing workflow integration from companies doing activity theater. (<a href="https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf?utm_source=chatgpt.com">MLQ</a>)</p><h4>3. The macro numbers do not yet scream AI productivity boom</h4><p>U.S. productivity growth slowed to an annualized 0.8% in Q1 2026, down from a revised 1.6% in the prior quarter and far below the 5.2% surge in Q3 2025. Productivity was still up 2.9% year-over-year, but the data does not yet show a broad, clean AI-driven productivity step change. (<a href="https://www.reuters.com/business/world-at-work/us-productivity-slows-further-first-quarter-2026-05-07/?utm_source=chatgpt.com">Reuters</a>) But those gains are coming almost exclusively from the infrastructure build out.</p><p>Stanford’s 2026 AI Index says organizational AI adoption continued rising, with AI used in at least one business function at 70% of surveyed organizations, while agent deployment remained in the single digits across nearly all business functions. Translation: AI is everywhere in adoption language, but true agentic operational deployment is still early. (<a href="https://hai.stanford.edu/ai-index/2026-ai-index-report/economy?utm_source=chatgpt.com">Stanford HAI</a>)</p><h4>4. Token economics are structurally different from SaaS economics</h4><p>Frontier APIs price intelligence as metered usage. OpenAI’s published API pricing lists GPT-5.5 at $2.50 per 1M input tokens / $15 per 1M output tokens for short context, with GPT-5.5 Pro at $15 / $90. Priority processing is materially higher. (<a href="https://developers.openai.com/api/docs/pricing?utm_source=chatgpt.com">OpenAI Developers</a>)</p><p>Open-weight and hosted open-model economics are much cheaper at the token level. DeepInfra lists Llama 3.3 70B Instruct Turbo at $0.10 input / $0.32 output per 1M tokens. Groq lists Llama 3.3 70B Versatile pricing around $0.29 input / $0.59 output per 1M tokens with high throughput. (<a href="https://deepinfra.com/pricing?utm_source=chatgpt.com">DeepInfra</a>)</p><p>The enterprise implication is obvious: frontier tokens may be worth it for hard reasoning, but high-volume bounded workflows will be aggressively routed to cheaper models, open weights, caching, deterministic software, and self-hosted capacity once CFO scrutiny arrives.</p><h4>5. The right metric is not cost per token</h4><p>The right metric is cost per verified business outcome.</p><p>A token can be:</p><ul><li>Cheap, when it becomes a verified business outcome.</li><li>Expensive, when it creates a draft that increases review burden.</li><li>Valuable, when it becomes code that passes tests, updates the system, and ships safely.</li><li>Waste, when it becomes another artifact in the corporate slop pile.</li></ul><p>This is the core economic mismatch. Frontier labs sell tokens. Enterprises need outcomes. Those are not the same unit.</p><h4>6. Insurance is already pricing AI as a new liability surface</h4><p>Verisk / ISO introduced optional commercial general liability endorsements for GenAI exposures, including CG 40 47 and CG 40 48, with exclusions for bodily injury, property damage, and/or personal and advertising injury arising from generative AI. (<a href="https://www.independentagent.com/vu_resource/verisk-to-roll-out-new-general-liability-exclusions-for-generative-ai-exposures/?utm_source=chatgpt.com">IndependentAgent.com</a>)</p><p>The Financial Times reported that insurers including AIG, Great American, and WR Berkley sought approval to exclude AI-related liabilities from corporate policies. One WR Berkley exclusion would bar claims involving “any actual or alleged use” of AI. (<a href="https://www.ft.com/content/abfe9741-f438-4ed6-a673-075ec177dc62?utm_source=chatgpt.com">Financial Times</a>)</p><p>This supports the risk thesis: underwriters are not treating GenAI as ordinary software risk. They are carving it out, limiting it, or forcing it into narrower risk corridors.</p><h4>7. Error rates remain too high for many business workflows</h4><p>The updated <a href="https://www.kaggle.com/benchmarks/google/facts/leaderboard">FACTS Benchmark Suite</a> shows frontier factuality still has a long way to go. As of the April 30, 2026 leaderboard, GPT-5.5 led with a 71.2% average score, implying roughly a 28.8% factual error rate across the benchmark. Gemini 3.1 Pro Preview followed at 67.7%, Gemini 3 Flash Preview at 63.4%, Gemini 2.5 Pro at 62.1%, and GPT-5 at 61.8%.</p><p>The category scores show why enterprise use cases require controls. GPT-5.5 scored 75.9% on grounding, 85.0% on search, and only 45.8% on multimodal factuality. Gemini 3.1 Pro Preview scored 85.6% on search, 65.0% on grounding, and 41.3% on multimodal factuality. Reliability varies sharply by task type, tool use, modality, and context quality.</p><p>FACTS measures four factuality modes:</p><ul><li><strong>Parametric:</strong> internal knowledge on factoid questions.</li><li><strong>Search:</strong> retrieval and synthesis using search.</li><li><strong>Multimodal:</strong> factual answers about images.</li><li><strong>Grounding:</strong> answers grounded in supplied context.</li></ul><p><strong>AI error rates range from low double digits in narrow, well-grounded, verifiable workflows to 25–40%+ across broad factual benchmarks, and much higher in specialized domains like law, medicine, finance, and compliance.</strong> Enterprise AI needs retrieval, validation, deterministic checks, audit trails, and human accountability before it can be trusted in material business workflows.</p><h4>8. Coding is the strongest domain because it is verifiable</h4><p>Programming has a unique advantage: errors can be detected by compilers, tests, type systems, linters, CI, AST checks, and runtime validation. That makes AI coding more economically tractable than most knowledge work.</p><p>The <a href="https://medium.com/brainly/from-vibe-coding-to-guided-coding-d2ba7e526ff3">Brainly / CodeStrap</a> study is the proof point. Larry AI achieved:</p><ul><li>80–90% one-shot correctness</li><li>1–2 turns total</li><li>~96–98% inference cost reduction</li><li>$0.04–$0.07 per task for Larry versus $1.78–$3.18 for Cursor/Kiro</li><li>Junior and mid-level engineers able to complete tasks that previously required deeper DAO expertise</li></ul><p>Larry won by encoding senior engineering knowledge into guided workflows, constraints, state-based execution, isolated worktrees/containers, and platform context. <strong>Not all domains will be able to repeat this performance because they are not verifiable.</strong></p><h4>9. The enterprise architecture likely shifts toward neurosymbolic systems</h4><p>Mature AI architectures will need to be cost effective. Cost effective AI systems that produce durable outcomes will need harness engineering that includes:</p><ul><li>State machines to control workflow.</li><li>Schemas to constrain inputs and outputs.</li><li>Retrieval to provide domain context.</li><li>Tests to verify correctness.</li><li>Determinism</li><li>Policy engines to enforce rules.</li><li>Ontologies to encode business objects and relationships.</li><li>Evaluators to measure pass/fail.</li><li>Model routing to reserve frontier models for hard cases.</li><li>Caching to avoid recomputing known outputs.</li><li>Open weights or self-hosting for high-volume bounded work.</li></ul><p>This moves intelligence out of the model and into the system around the model. That reduces cost, variance, review burden, and dependency on frontier API pricing. Recent studies including <a href="https://arxiv.org/abs/2603.28052">one from Stanford</a> show the harness around the model is far more important than the model itself (6x more important).</p><h4>10. The IT-spend backtest says AI does not disappear, but bad spend gets repriced</h4><p>Enterprise IT went through its own ROI reckoning in the early to mid 2000s. Computers, ERP, CRM, cloud, analytics, and cybersecurity did not always show clean audited ROI at first, but the spend survived because it became the operating substrate of the enterprise. When the CFO got involved we had to measure the ROI and the outcome. As a result, we got agile and we got SaaS.</p><p>The lesson for AI:</p><ul><li>Spend survives when it becomes a durable systems.</li><li>Spend gets cut when it remains an activity.</li><li>Software compounds; token calls usually do not.</li><li>Systems of record persist; generated drafts usually do not.</li><li>Workflow state persists; chat output often does not.</li></ul><blockquote><strong>The likely future is not lower AI usage. It is more disciplined AI usage.</strong></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1f165de8cddd" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codestrap/ten-facts-and-findings-on-enterprise-ai-spend-tokenomics-and-roi-1f165de8cddd">Ten Facts and Findings on Enterprise AI Spend, Tokenomics, and ROI</a> was originally published in <a href="https://medium.com/codestrap">Codestrap</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>