The multi-AI orchestration CLI — engines compete to build, collaborate to think.
Agon AI puts the world's best AI engines on your software engineering problems two ways. They compete on implementation: the same task runs in isolated git worktrees, the best solution wins and is applied automatically. They collaborate on judgment: brainstorm panels feeding a synthesis, council roundtables with assigned roles and a chair, tribunal debates, and multi-engine review consensus — pooling model strengths instead of betting on one. Glicko-2 ratings continuously track each model's performance over time, so routing keeps getting sharper.
npm install -g @kernlang/agon- What is Agon?
- Requirements
- Installation
- Core Modes
- Interactive REPL
- Rooms
- Using Agon from Other CLIs
- Engines
- Ratings
- Cesar Routing
- Configuration
- Mods
- Updating
- Architecture
- License
The agon REPL on launch — your roster, and plain prompts routed to the right mode.
Agon AI fundamentally changes how developers interact with AI coding assistants. Instead of relying on a single model and hoping for the best, Agon introduces competitive orchestration. You describe the task once, and Agon dispatches it to multiple models simultaneously. They race to implement the solution, and you review and apply the best outcome.
Under the hood, Agon uses a sophisticated worktree isolation model. When a competition begins, Agon creates temporary, hidden git worktrees for each participating engine. This ensures that engines can modify files, run tests, and iterate without stepping on each other's toes or cluttering your main working directory. Only the winning implementation is merged back into your active branch.
Routing these tasks efficiently is Cesar, Agon's intelligent orchestration brain. Cesar monitors every interaction, success, and failure, updating an internal Glicko-2 rating system for each engine. Over time, Cesar learns which engines excel at specific task classes—like writing tests, refactoring React components, or optimizing database queries—and automatically routes future tasks to the highest-rated engine for that specific domain.
Agon isn't just about competition; it's a unified platform for AI collaboration. Whether you need engines to debate a controversial architectural decision, brainstorm solutions with weighted token allocations based on confidence, or simply execute a robust pipeline loop, Agon provides the ultimate terminal-native interface for multi-agent workflows.
- Node.js >= 22
- A Git repository
- At least one supported AI CLI installed globally (e.g., Claude Code, Codex, Antigravity) or an API key for API-based engines
- Optional: Python 3.10+ to unlock the semantic features that run as Python sidecars — semantic history search (
agon history --query), tree-sitter syntax validation in forge fitness, brainstorm paraphrase dedup, and the task classifier. Agon still runs without Python; these features fall back to substring/regex paths.
npm install -g @kernlang/agonThat's it — one command, no clone, no build. It pulls the engine substrate
(@kernlang/agon-engines) and the semantic sidecars (@kernlang/agon-dedup)
in automatically. Then run agon in any git repository to start the
interactive REPL.
# (Optional) Install the Python deps to unlock the semantic features
# (semantic history search, tree-sitter fitness, dedup, task classifier).
# The .py sidecars ship with the package; this just installs their libs.
python3 -m pip install --user fastembed numpy tree-sitter tree-sitter-python tree-sitter-typescript tree-sitter-javascript tree-sitter-json# 1. Clone the repository
git clone https://github.com/KERNlang/agon.git
cd agon
# 2. Install dependencies (pulls @kernlang/agon-engines + the Python sidecars)
npm install
# 3. Build and install the global CLI
npm run install:cli
# 4. (Optional) Install Python sidecars for semantic features
python3 -m pip install --user -r packages/dedup/requirements.txtOnce installed, run agon in any git repository to start the interactive REPL. Run agon doctor to verify every engine, the worktree path, and the Python sidecars resolve.
Agon features multiple operational modes to suit different development challenges.
Pick by the shape of the problem, not the topic:
| You need… | Use | Why |
|---|---|---|
| Ideas / options / "what am I missing?" | brainstorm | Cheap and fast — engines bid confidence and surface approaches and gaps. |
| One refined output from many opinions | synthesis | Engines draft, improve each other's drafts over swap rounds, and a judge picks the best evolved result. |
| A decision with real tradeoffs settled | tribunal | Structured debate (adversarial / steelman / socratic / red-team / synthesis / postmortem) that argues the sides. |
| The whole panel on a high-stakes call | council | A roundtable — every engine takes a distinct role (Contrarian = top critic, Red-Team, …); a top-rated chair returns ONE verdict with a confidence + kill-switch. |
| Open exploration, no decision yet | campfire | Relaxed multi-engine discussion — no scoring, no winner. |
| A problem decomposed before you act | think | Sequential thinking — linear or reflexion (forced self-critique), optional branches; surfaces open questions + a goal handoff. |
| Your own decision pressure-tested | nero | Adversarial self-challenge — the top-rated critic (by tribunal rating, not the best builder) attacks it and returns a verdict. |
| A library/repo/spec/fact looked up with sources | research | Keyless, web-grounded, cited research — Agon discovers sources (npm/GitHub/MDN/IETF/Stack Overflow/Wikipedia, no API key), an engine drafts a grounded answer, and Agon verifies every citation. |
| Your own browser driven to research / check a page | chrome | Drive your real browser (navigate / read / screenshot / click) from the REPL, the CLI, or Cesar — reuses a running agon the side panel is on, else embeds a transient bridge. Needs the Agon browser extension. |
| A node-graph flow checked before anything runs | flow |
