// the find
coleam00/Archon
The first open-source harness builder for AI coding. Make AI coding deterministic and repeatable.
Archon is a YAML-based workflow engine that wraps AI coding agents (primarily Claude Code) in deterministic pipelines — plan, implement, validate, PR — with git worktree isolation per run. It's for teams or solo developers who want repeatable AI-assisted development processes rather than one-shot prompts. Think of it as a Makefile for AI coding tasks, with human approval gates and parallel execution.
The git worktree isolation is the right call — running 5 parallel AI tasks without branch collisions is a real problem this solves cleanly. The mix of deterministic nodes (bash scripts, git ops) and AI nodes is architecturally sound; you're not forcing AI into steps where it adds no value. The human-in-the-loop approval gate (`until: APPROVED`) is honest about where you still need a human rather than pretending full automation. The security scanning for marketplace workflows (fixtures for RCE, reverse shells, credential leaks) shows they've thought about the trust model.
Hard dependency on Claude Code as the AI runtime is a significant lock-in — despite mentioning Codex and Pi support, the entire setup wizard and skill system is Claude Code-first, so switching costs are real. The `fresh_context: true` loop pattern is papering over a fundamental problem: long-running AI context degrades, and restarting context each iteration loses cumulative understanding of what was already done. No mention of cost controls or budget caps on workflow runs — a looping AI node that hits `until: ALL_TASKS_COMPLETE` on a stubborn bug could burn through significant API spend with no guardrails. SQLite as the default persistence for a multi-workflow, potentially multi-user system will serialize everything and become a bottleneck; the PostgreSQL option exists but isn't the default.