Skip to main content

Multi-agent skill

Agent Orchestration Skill

Scan running agents, assess progress, and coordinate multi-agent workflows.

On this page

Install command

AI DevKit installs its built-in skills as one curated set. Run this once to install agent-orchestration with the rest of the built-in workflow skills.

bash
npx ai-devkit@latest skill add --built-in

What it does

Use agent-orchestration when several AI agents are working at once and someone needs to act like the technical lead. It helps check progress, identify blocked work, and send next instructions until the workflow is complete.

When to use it

  • Manage multiple active AI coding agents
  • Check progress and unblock parallel work
  • Coordinate multi-agent implementation, review, or research tasks

Supported agents

  • Claude Code
  • Codex
  • Gemini CLI
  • opencode
  • Pi

The SKILL.md file

This is what your agent reads when the skill runs. View it on GitHub (opens in a new tab)

Use only for multi-agent supervision: coordinating dependencies, polling progress, unblocking waiting agents, relaying outputs, resolving conflicts, and verifying completion across agents. For one-off list/detail/send/start/kill work, use agent-management or agent-communication.

Use agent-management for safe agent selection and lifecycle actions. Use agent-communication for list/detail/send mechanics. Use verify before accepting any agent's completion claim.

Rules#

  • Own the loop until assigned work is complete, blocked, or stopped.
  • Run agent list --json before each pass; never assume names/statuses.
  • Inspect waiting, idle, unknown, missing, or stale agents before acting.
  • Send self-contained instructions and avoid duplicate follow-ups.
  • Sequence agents that touch the same files; relay only relevant upstream output.
  • Escalate only for repeated failures, unresolved conflicts, product/business decisions, or destructive/shared/production/security-sensitive actions.

Loop#

If the goal or agent ownership is unclear, run one scan/detail pass. Ask the user once only if context is still insufficient.

  1. Scan agents.
  2. Assess agents needing attention with detail --tail 10.
  3. Act: approve, clarify, correct, delegate, relay, verify, or escalate.
  4. Report one brief status line.
  5. Run ai-devkit agent wait --json (add --id/--group to watch only your agents), then repeat. It blocks until an agent moves to waiting, idle, or exited, or until its 5-minute default timeout; reason: "timeout" is normal, so do a light scan and wait again. If your command timeout is shorter, run it in the background or raise that timeout. Never use a long sleep: harnesses block or kill it.

Completion#

Finish when all assigned work is verified, blocked with a clear reason, or stopped by the user. Summarize per-agent outcomes, verification, unresolved issues, and next step.

esc