What it does
Agent Coordinator is a skill you install for your own user account and invoke in Codex with a $coordinator prefix. It is meant for work with several dependent steps, files or specialties. The page says it is not needed for one small, obvious step. For each job it follows four steps: make the outcome and the proof of success explicit, split the work into small parts with clear boundaries and dependencies, work through the ready parts in a safe order, then inspect the evidence before finishing. Each part gets acceptance criteria, one role and declared write scopes, and independent parts run side by side only when their scopes do not overlap. Progress is recorded under ~/.agent-coordinator, so a new run can resume after an interruption, mark work that may still be active for reconciliation, and avoid starting an uncertain step twice. Specialist agents are optional; without them each part is done inline. Completion reruns the proof commands recorded for the plan. It needs Python 3.11 or newer and no third-party packages. Installation is a single prompt that follows the repository's INSTALL.md and does not change Codex settings. A Docker demo is included and needs an OpenAI API key.
Capabilities
- Splits a request into small parts with dependencies, owners and acceptance criteria
- Records progress in a private folder so interrupted work can resume
- Reconciles uncertain work before retrying it
- Reruns proof commands before a workflow is marked complete
- Works with optional specialist agents or runs everything inline
- Installs for the current user without editing Codex settings
When to use it
- A job has several dependent steps, files or specialties
- Independent pieces could safely move ahead in parallel
- An interruption would make progress hard to reconstruct
