Runbooks the agent actually follows
DevOps, SRE and securityOps team with runbooks in the repo
Our runbooks live in docs/runbooks. When we asked an agent to rotate the Redis TLS certificate, it improvised instead of opening the runbook, skipped the 'verify replicas' step, and one replica kept the old certificate for a day.
How do I make an AI agent follow our runbooks step by step instead of improvising?
Add an ArcRoom rule on your runbooks folder, like 'follow these step by step for any ops task', and recall shows it when a prompt mentions rotation, failover or restore. Add a read-only verify script as a done-check in .arcroom.json, approved once with arcroom checks approve. The agent can't stop until it passes. ArcRoom doesn't enforce each runbook step.
Also asked as: Our agent skipped a runbook step and left a replica on an old certificate. How do we prevent that? · How do I get an AI agent to use the runbooks stored in our repo? · Can I stop an agent from finishing until a verify script passes?
A rule marks the folder: 'docs/runbooks: follow these step by step for any ops task'.
- Done-checks
- Recall
Done-checks (.arcroom.json): ./scripts/verify-redis-tls.sh ✗
Step by step
- A rule marks the folder: 'docs/runbooks: follow these step by step for any ops task'.
- Recall puts that note in front of the agent when the prompt mentions rotation, failover or restore.
- A read-only verify script is added as a done-check in .arcroom.json, approved once with arcroom checks approve.
- The agent can't stop until the verify check passes.
What the agent sees
Done-checks (.arcroom.json): ./scripts/verify-redis-tls.sh ✗
replica-2: certificate expires 2026-11-02 (old)
Not done: finish the runbook step 'verify replicas'.Compared with today's tools
CLAUDE.md pointer to the runbooks
Add 'always follow docs/runbooks' to CLAUDE.md and hope the agent opens the right file.
- Setup
- 10 min
- Ongoing
- remind the agent in the prompt each time
ArcRoom
Every agent reads the team playbook before it works and writes to it when it finishes. Nothing to update by hand.
- Setup
- 2 min plus one check approval
- Ongoing
- None
CLAUDE.md can say 'verify', but nothing stops the agent from finishing early; the done-check does.
Where it stops
ArcRoom doesn't enforce each runbook step, only the done-check at the end. A done-check should be read-only; it runs every time the agent tries to stop.
Try it in your project
Works with the agents your team already uses, and any agent over MCP or the HTTP API.