What should an agent save so a failed run can be replayed safely?

A failure trace helps explain why an agent failed. Recovering safely takes more context. As I’m implementing recovery for agent workflows, I’m separating two scenarios: • Reproduce the failure: preserve the inputs, prompt and tool versions, tool responses, and approval decisions. • Resume after a fix: recover from a safe checkpoint and check which external actions have already happened. A retry shouldn’t send the same message twice or create a duplicate record. Tracking external IDs and checking for completed actions are part of how I’m handling this. I’m also balancing recovery with retention: keeping enough context to understand decisions without storing every sensitive payload. That record also shapes the operational response when a dependency fails: retry automatically from the checkpoint, notify someone with the relevant failure trace, or wait for a person to fix the issue and approve the rerun. What’s the minimum an agent needs to save to make an automatic retry safe? And when should a person step in?