The Fleet Runbook: Operating an Agent Swarm by the Book
An agent fleet that runs on tribal knowledge is a fleet that fails in the dark. The runbook is the operations artifact that turns a swarm of agents into an operable system: the documented procedure for every routine action — deploy, restart, verify, scale, decommission — written down before it is needed. The fleet runbook is the operational translation of everything the research track established (BR-223): evidence becomes instruction.
What a fleet runbook contains
The runbook has fixed sections per procedure: trigger (what condition starts this), preconditions (what must be true before acting), steps (the exact commands and checks), expected result (what success looks like), and rollback (how to undo). Procedures cover the full lifecycle: onboarding a new agent profile, deploying a skill update, restarting a crashed worker under its supervisor (S1.3), rotating a credential (S7.5), and decommissioning a retired agent. Each procedure carries a pointer back to its evidence — the research or post-mortem that justified it.
Runbooks as the audit layer
The runbook is also the audit artifact: an operation performed from a runbook is a verifiable operation. The security series taught that trust requires evidence (S7.7, S7.15); the runbook is where that evidence lives for operations. When an incident occurs, the runbook is both the playbook and the record — the operator follows the documented steps, logs what happened, and the log becomes part of the post-mortem (OP-10).
Runbooks live, or they rot
A runbook that is never exercised rots: commands drift, steps go stale, and the first time it is needed it fails. The discipline is to exercise runbooks on a cadence — the same cadence as the deployment schedule (OP-2) — and to update them whenever a procedure changes. The fleet runbook is a living document, versioned like code, and it is the first thing a new operator reads. The fleet is only as operable as its runbook is current.
Grounded in the OP operations series, the S7 security runbooks (S7.11), the S1 supervision-tree work, and the BR-223 evidence-to-runbook bridge. First article in the Round D operations track.
