Is that agent actually ready for production?
Most agent failures are not model failures — they are governance failures that were visible before launch. Verify these nine conditions before any autonomous agent touches production systems.
Shared human credentials make agent actions indistinguishable from user actions and destroy accountability.
Define exactly what the agent may do, for how long, and how that authority is withdrawn — architecturally, not by policy document.
Decide which action classes (payments, deletions, external sends, deploys) require human sign-off before the agent exists, not after.
If the agent went fully wrong for one hour, what could it reach? If the answer is “unknown,” it is not production-ready.
Not “we would revoke the API key” — a tested, timed procedure that halts the agent mid-run.
Recovery means restoring the systems the agent touched to a known-good state, not just stopping the agent.
Session, tool, arguments, result, timestamp. If a regulator asked for the trail tomorrow, you could produce it.
Alignment at launch is not alignment in month three. Baseline behavior and alert on deviation.
A named human accountable for the agent’s actions, reviews, and retirement. No owner, no launch.