multi-agent-coordinator

When running 3+ OpenClaw agents in parallel, the flat agents.list[] config becomes unmanageable, channel "bleeding" occurs between agents, parallel agents can produce contradictory outputs, and there is no timeout detection for silent agent failures.

This skill lives in the orchestrator agent and provides: a shared agent registry, health-check heartbeats, output consistency checks, and structured inter-agent handoffs.

Agent registry

python3 run.py --register agent-id=coder role=code_implementation channel=C001
python3 run.py --register agent-id=reviewer role=code_review channel=C002

The registry tracks: agent ID, role, channel, status, last-seen timestamp, and current task.

Health checks

The orchestrator pings each registered agent's last-seen timestamp. An agent is considered timed out if it hasn't updated in longer than the configured timeout (default: 30 minutes).

Invoke periodically:

python3 run.py --health-check

On timeout detection: write AGENT_TIMEOUT to state, alert the orchestrator, and prompt: "Agent coder hasn't responded in 45 minutes. Reassign task or wait?"

Output consistency checkpoint

Before merging parallel agent outputs:

python3 run.py --consistency-check --agents coder reviewer --key "api_design"

The orchestrator compares outputs for the same key and flags contradictions. Example: coder proposes POST /users, reviewer proposes PUT /users/{id} for the same operation — these are flagged as contradictions requiring human resolution before merge.

Structured handoff

Pass context from one agent to another:

python3 run.py --handoff --from coder --to reviewer --task-file handoff.yaml

Handoff files capture: what was done, what wasn't, key decisions made, constraints discovered, and next recommended action. This extends task-handoff (which is single-agent) to the multi-agent case.

Difference from workflow-orchestration

workflow-orchestration executes a linear sequence of steps in a single agent. multi-agent-coordinator manages a parallel fleet: independent agents with their own sessions, channels, and context windows.