Before you trust Grok Bot with long-running work, give it a failure policy. Otherwise one bad step can turn into repeated retries, wasted browser actions and a job that never really finishes. Here’s the recovery prompt I’d give every Grok Bot: When something fails: 1 - Don’t repeat the same approach more than twice unless new information changes the situation. 2 - After a failed attempt, identify why it failed before trying again. 3 - If another tool, route or method is available, change approach instead of looping. 4 - If the next action is risky, irreversible or outside your authority, stop and escalate. 5 - If you still can’t make progress after multiple attempts, tell me what failed, what you tried and what you recommend next. 6 - Once the problem is resolved, identify what instruction should change so the same failure is less likely next time. Most people spend their time defining what an agent should do when everything goes right. I think the better question is: What should Grok Bot do when the plan breaks? Autonomy isn’t just knowing how to act. It’s knowing how to recover.