Swarm Coordinator

# Swarm Coordinator

Built by the Dead Brands team (deadbrands.co).

Outside helper: Rock.

Run a multi-agent crew as a team: clear hierarchy, heartbeats with thresholds, shared state, and calm takeovers when a lane goes silent.

## Purpose

Teach an outside operator how to coordinate many agents without silos, stuck approvals, or silent failures — so work keeps moving and evidence flows up.

## When to use

- You have three or more agents sharing one product or ops lane.
- A lead went quiet and work is stalling on an unclicked prompt.
- Two agents keep rediscovering the same facts because state lives in one head.
- You need a takeover rule that is mechanical, not personal.

## Limits

- This skill does not grant authority over other people's machines or accounts.
- It does not invent monitoring tools. Use the heartbeats and logs you already have, or mark new ones "Planned, not yet available:".
- Never encode secrets, private hostnames, or personal blame into the shared state.
- Reassignment notes the work, not the worker.

## Steps

1. Draw the ladder: owner → lead → managers → workers. Write who acts when the lead is silent.
2. Define a heartbeat cadence and a silence threshold (example: beat every 15 minutes; act after 45 minutes of silence).
3. On silence: check for a stuck human approval gate first; then reassign outstanding work; leave a note the original lane will see.
4. Keep one shared changes log and one runbook set every lane can read.
5. Hand results up as evidence (links, receipts, hashes), never as bare claims.
6. When the original lane returns, give them a fresh task — no blame ritual.

## Examples

Input: Lane B missed three heartbeats; a draft is half-done.

Output:
- Check: approval card still open for the owner → wait, ping once.
- If no gate: reassign draft to Lane C with the current file hash and next step.
- Log: "2026-10-28T11:00Z reassigned draft X to Lane C after 45m silence; note left for Lane B."