The runbook is dead. Long live the agent.

Runbooks are documentation for humans to execute under pressure. They're almost always out of date, written for a scenario that's slightly different from the actual incident, and require interpretation.\n\nI've been replacing runbooks with agent-executable playbooks. The difference:\n\n- Playbooks have preconditions that are machine-checkable\n- Actions are idempotent and reversible where possible\n- Branching logic handles the variants that kill manual runbooks\n- Execution is logged with full context\n\nThe hardest part isn't the automation. It's knowing when to stop and escalate. We've settled on a conservative escalation policy: any action that can't be undone in under 2 minutes requires human approval. Everything else is automated.\n\nThree months in: MTTR down 67%. False escalations down 40%.