Ask AI to fix a bug, and three runs might yield three different results.
Sometimes it forgets to run tests, sometimes the PR description is half-baked, sometimes it skips planning and dives straight into code — every run depends on the model’s “mood.” The biggest pain point of AI coding tools isn’t lack of capability, but lack of determinism.
Archon (coleam00/Archon) takes a direct approach: Dockerfile standardized infrastructure, GitHub Actions standardized CI/CD, so AI coding workflows need standardization too.
From Agent Builder to Workflow Engine
This project underwent a complete rewrite — migrating from Python to TypeScript, and shifting its positioning from “AI Agent Builder” to “AI Coding Workflow Engine.” The core change: instead of trying to let AI autonomously decide what to do, the development process is encoded as YAML workflows, with AI stepping in only where intelligent judgment is needed.
A typical workflow looks like this:
nodes:
- id: plan
prompt: "Explore the codebase and create an implementation plan"
- id: implement
depends_on: [plan]
loop:
prompt: "Read the plan. Implement the next task. Run validation."
until: ALL_TASKS_COMPLETE
- id: run-tests
depends_on: [implement]
bash: "bun run validate" # Deterministic step, no AI needed
- id: review
depends_on: [run-tests]
prompt: "Review all changes against the plan."
- id: create-pr
depends_on: [review]
prompt: "Push changes and create a pull request"Five nodes — plan → implement → test → review → PR. Only implement and review require AI; run-tests is purely deterministic. AI is no longer a “full agent” but an “on-demand intelligence module.”
17 Preset Workflows
Archon ships with 17 workflows covering the vast majority of daily development scenarios:
| Workflow | Purpose |
|---|---|
| archon-fix-github-issue | Triage → Investigate → Implement → Verify → PR → Review → Self-Heal |
| archon-idea-to-pr | Idea → Plan → Implement → Verify → PR → 5 parallel reviews → Self-Heal |
| archon-comprehensive-pr-review | 5 parallel reviewers with automatic fix |
| archon-smart-pr-review | Complexity-classified, targeted reviews |
| archon-architect | Architecture scanning to reduce complexity |
| archon-refactor-safely | Safe refactoring with type-check hooks |
| archon-resolve-conflicts | Detect → Analyze → Resolve → Verify conflicts |
| archon-plan-to-pr | Execute existing plan through to PR |
You can also define your own YAML workflows in the .archon/workflows/ directory — once committed to the repo, the entire team runs the same process.
Key Design: git worktree Isolation
Each workflow run executes in an isolated git worktree. This means you can kick off 5 fix tasks simultaneously with no conflicts, each working on its own branch. When done, you return directly to a branch with a PR already created.
“Fire and forget” — launch it and walk away, come back to collect the PR.
Multi-Channel Triggering
Beyond CLI and Web UI, Archon supports remote workflow triggering from Slack, Telegram, Discord, and GitHub Webhooks. Telegram and Discord each take about 5 minutes to set up; Slack and GitHub take roughly 15 minutes.
The Web UI includes a chat interface, real-time workflow monitoring dashboard, drag-and-drop workflow editor, and a unified message center — messages from all platforms converge in one place.
Architecture & Dependencies
Platform Adapters (Web/CLI/Telegram/Slack/Discord/GitHub)
↓
Orchestrator
↓ ↓
Command Workflow AI Assistants (Claude / Codex / Pi)
Handler ExecutorThe storage layer is SQLite or PostgreSQL (7 tables), managing codebases, conversations, sessions, and workflow run records.
Currently primarily relies on Claude Code (also supports Codex and Pi). Installation options include Homebrew, one-click script, and Docker. Building from source requires Bun + Claude Code.
Stats
- ⭐ 20.5k Stars, 3.1k Forks
- 1,355 commits, latest commit 9 hours ago
- Written in TypeScript, active development branch
- MIT License
- Website: archon.diy
Archon isn’t another “code-writing Agent.” It tackles a deeper problem: once you already have an Agent, how do you make it produce the same result every time. For teams worn down by the unpredictability of AI coding, this project is worth watching.