Your agent returned 200 OK and did the wrong thing

Ordinary software tells you when it breaks. AI systems return well-formed, confident, wrong answers with a success status. That one property is why observability for AI is a different discipline rather than the old one pointed at new infrastructure.