An agent in production is a system like any other. It needs monitoring, an on-call rota and a written answer to “what do we do when it is wrong?”
What goes in the runbook
How to tell the agent is misbehaving, who is paged, how to switch to the manual path, and how to replay the work once it is fixed. One page per failure mode is enough.
Costs are part of operations
Track cost per task next to accuracy. A workflow that is right ninety-nine times in a hundred and three times more expensive than the manual process is not finished.
“If nobody is on call for it, it is still a demo.”
Key takeaways
Write one page per failure mode.
Keep a manual path you can switch to.
Track cost per task beside accuracy.
Lena Fischer
Partner, Automation. Takes agents and workflows from prototype to monitored production, with the runbooks to keep them there.
