If you’re letting an AI agent touch production, a green unit test is not enough. The failure mode is rarely “the model was dumb” — it’s usually missing guardrails around scope, rollback, credentials, or observability. Here’s the lightweight checklist I’d want before any agent can run outside a sandbox:
Before execution, force the agent to say:
If a human cannot read the plan in 30 seconds, the task is still too vague. A lot of teams mix safe diagnosis with risky execution.
Keep them split:
That one boundary removes a surprising amount of risk.
You want both:
Example: “If checkout latency rises or 5xx errors spike after deploy, revert immediately.”
For every production-touching run, capture: Without that trail, postmortems become guesswork.
“Nothing broke” is not the same as “it worked.”
If you cannot observe the key event, you cannot close the loop. At minimum, define one thing you expect to move after the change.
One thing I keep seeing: teams want a full “agent governance platform” before they’ve done five manual audits.
Usually the faster move is:
That gives you better language, clearer scope, and fewer fake features.
I’m experimenting with a small Agent Readiness / Governance Audit workflow around exactly this problem. Not a full platform claim — more a practical preflight for teams that want to catch the obvious mistakes before an agent touches production.
Curious where your own process still feels weakest today: