Swarm Coordination

Swarm coordination (3+ agents)

Two-agent coordination is mostly handoffs. Three-or-more-agent coordination has new failure modes: silent clashes, redundant work, and consensus collapse. Use this skill when the orchestrator (you) is fan-ing out to 3+ workers on the same objective.

When to use

• Sprint planning where multiple agents work the same backlog.
• Multi-repo refactor where each repo has its own agent.
• Research where multiple agents explore different angles of the same question.

The shape

• One orchestrator. That is you. No two orchestrators on the same swarm.
• N workers. Each has a clearly named lane (e.g., worker-backend, worker-frontend, worker-research).
• One shared scratchpad. A planning task description, a doc, or a thread that all workers and the orchestrator can read. No private channels.
• One arbitration rule. When workers conflict, the orchestrator decides; if the conflict is policy-shaped, escalate via consensus-arbitration.

The loop

1. Roster. Confirm each agent is clocked in via agent_directory_list.
2. Lane assignment. Write each lane's scope to the shared scratchpad. Each worker confirms (status: starting) on its lane line.
3. Heartbeat. Every worker runs agent-status-pings at 3-minute cadence.
4. Conflict checks. Every 10 minutes, the orchestrator scans the scratchpad for overlaps and either splits/merges lanes or invokes consensus-arbitration.
5. Convergence. When all lanes are done, the orchestrator runs production-validation (if engineering) or agent-self-check (if non-engineering) on the unified output before reporting back to the user.

Anti-patterns

• No shared scratchpad — workers chat in private threads and drift.
• Two orchestrators reassigning the same lane.
• Letting a blocked worker stall the swarm — re-route the lane within 1 cycle.