Define who owns a failure in AI-written code

Your org wants a rule for who is accountable when AI-written code fails in production. Today it is fuzzy. An engineer ran the agent, skimmed the diff and merged, and when the code broke, the incident review drifted into "the AI wrote it".

As the staff engineer, what ownership and merge model do you set: where does accountability sit, what must review mean for agent code, and which perverse incentive do you design against?