Make agent-written changes explainable after an incident

Your team merges many agent-written PRs. A pricing bug shipped, and in the postmortem someone asked why the discount code was written that way. The author re-ran the agent and got a different implementation. The original prompt was gone, and nobody could say what the reviewer had checked.

What process do you put in place so the next postmortem can answer that question? Decide what gets recorded, who owns it, and what "reproducible" can and cannot mean with a model.