What Is an Agent Runbook?
A specification for what an agent should produce, how it may work, and when it should return.
An execution agreement
An agent runbook turns an open-ended request into a bounded job. It defines the desired outcome, relevant context, scope, constraints, permissions, approval gates, recovery behavior, and acceptance criteria. The result is a prompt you can paste into your chosen agent.
Start with the deliverable
“Research competitors” describes an activity. “Produce a sourced comparison of five relevant competitors, with buyer tradeoffs and three strategic implications” describes an assessable outcome. State what should exist when the run ends.
Use only the sections the job needs
A small task may need little context and no custom strategy. A substantial project benefits from explicit boundaries and recovery instructions. Leave out empty optional sections instead of padding the runbook.
Instructions have limits
A runbook guides behavior; it does not enforce permissions, guarantee correctness, or grant tools the agent does not have. Review consequential results and use your agent environment’s access controls.