Everyone keeps saying AI agents can run your operations now. So I gave one seven days to try, on a real environment, not a demo. Here's what actually happened, including the part that nearly went wrong.
The setup
I connected an AI agent to my monitoring stack and cluster with read access first, then limited write access for a few safe actions like restarting a stuck pod or scaling a deployment. It could see logs, metrics, and recent deploys, and it could suggest or take a small set of pre-approved actions. Anything bigger needed my sign-off.
The rule I set for myself: treat it like a new junior engineer on their first on-call. Useful, fast, and absolutely not trusted with the keys yet.
What went well
