I wrote this after a painful production incident. An agent's reasoning depth dropped 67% between two model updates — zero error rates, HTTP 200 on every call, valid JSON throughout. We found out three weeks later from a customer complaint. No alert had fired.