Your runbook has two readers at 3AM now. The tired human it was written for, and the agent standing next to them. The human fills every gap with judgement. The agent fills it with action. "Restart the service if it looks unhealthy" was a fine instruction for one of them. For the other it's four unanswered questions and a production change. This talk is what we learnt putting agents in the incident channel and watching them run our runbooks exactly as written, which turned out to be the problem. Three real failure modes: a confidently hallucinated diagnosis, an unsafe action that free reasoning decided was reasonable, and "remove outdated entries" where nobody had ever defined outdated. Every adjective is a landmine. The fix isn't five hundred lines of YAML. It's making six decisions explicit that a senior engineer already makes in their head: trigger, precondition, verify, stop condition, rollback, and where a human must sit. We'll show the same runbook before and after, the existing schemas you should borrow rather than invent, how to climb from read-only to autonomous one rung at a time, and the part that finally fixes staleness: the agent proposing runbook edits as PRs after every incident, with a human still holding the merge button. The runbook isn't dying. It's being promoted from documentation to interface.
Navin is Director of Engineering at StackGen, where he leads the team building the future of Agentic Ops. You'll often find him on the internet stoking flamewars with his opinions about Observability, Open Source, and Dev Culture.