A meta-agent that orchestrates agents to fully automate computer use.

AgentChain (coming soon…) is a CLI (ac) and config system (~/.agentchain/) for running a hierarchy of AI agents automate computer use semi-efficiently.

the architecture #

Three tiers:

  • CTO: the meta-agent. Writes OKRs, launches projects with a GM, monitors progress, verifies deliverables. Orchestrates work.
  • GMs: one per project. Owns delivery against OKRs, breaks work into tasks, delegates to team agents
  • team agents: specialists (engineer, product, docs, etc.) that execute focused work

Each tier runs in its own tmux session. The CTO lives in cto, GMs in gm-<project>. State persists via NOTES.md, TODO.md, and a SQLite database for agentchain metadata (meta.db).

the config #

Configuration lives in ~/.agentchain/:

  • config.toml: user config, defaults, profiles (work vs. life)
  • CTO.md / GM.md: system prompts for each agent tier
  • teams/: agent definitions grouped by team
  • patterns/: prompt templates for different project types
  • project-templates/: repo lists and branch conventions
  • tasks/: recurring tasks (triaging email, GitHub issues, self-improving agentchain, etc.)

Profiles auto-resolve based on machine type. A work machine gets work defaults, a personal machine gets personal defaults.

the workflow #

I don’t run ac project create, the CTO does (though I still can). Usually, my interface is conversation through one UI or another. I tell the CTO what I want, and it handles the rest: writing the brief, launching GMs, monitoring progress, verifying deliverables, closing projects. I have a few skills (/up, /down, /status, /notes) for the basics, but the CTO is the one running ac commands.

/up                      # start the CTO
/status                  # dashboard: CTO, active GMs, recent events
/down                    # shut it all down

The CTO writes a brief with OKRs to briefs/foo.md, then runs ac project create – which creates a project directory, clones repos from a template, symlinks the right agents and patterns, and launches a GM in a new tmux session. The GM reads its OKRs, breaks them into tasks, delegates to team agents, and reports back. The CTO monitors via /loop, nudges if stuck, and verifies deliverables before closing.

why it matters #

It’s a more efficient way to use a computer – or will be!