A written procedure that tells an on-call engineer (or an AI agent) exactly how to respond to a specific kind of incident.
A runbook is a written, executable procedure for responding to a specific incident pattern, scaling pods on traffic spike, rotating an IAM key, flushing a stale cache, restoring a replica from snapshot. Modern runbooks live next to the alert that triggers them, are tested in game days, and are scored on whether they actually work when called. AI runbooks add machine-readable preconditions and actions so an agent can execute them inside a policy envelope without waking a human for routine cases.
Tribal knowledge runbooks (in someone's head, on a Confluence page from 2019) fail at the worst possible time, when the person who knows them is on vacation and the page is fired at 3am. Codified, version-controlled, executable runbooks turn institutional memory into a reliable artifact, and unlock agent-driven remediation as a downstream benefit.
See the part of the platform that handles runbook in production.