Archon setup: YAML workflows for Claude Code agents

Archon adds a defined workflow around an AI agent's coding tasks. The speaker calls this harness engineering: planning, coding, testing and review become steps in YAML, while reusable agent skills supply instructions automatically. The video contrasts this approach with the context drift and changing plans the speaker encounters in Claude Code, Cursor and Codex.

The demonstration runs Archon locally on an M4 Pro. The speaker starts its interface with Archon serve, installs an Archon skill into a repository using Claude, then launches an issue-fixing workflow. The agent discovers the skill and follows the workflow. The terminal and UI expose progress, logs and output; when a step fails, the UI shows where the workflow needs attention. The example ends with a pull request.

Archon's YAML DAGs combine AI steps with fixed steps. Each run uses a separate git worktree. The speaker credits this isolation with preventing agents from overwriting one another and claims consistent results without merge conflicts. These are the presenter's claims, rather than guarantees established by the demonstration. The comparison with LangChain focuses on Archon's emphasis on coding workflows.

The speaker describes Archon as open source but notes that workflow design takes effort, the project is still evolving and model choice affects output. For quick prompts, they suggest the extra setup may be unnecessary.