Engineer shares 4-field logging schema that exposes AI agent failures standard logs miss

A software engineer discovered that their production AI agent pipeline was silently hiding failures despite logging every prompt, tool call, and response — because the logged fields were structured for cost analysis rather than failure detection. The issue surfaced when an agent completed a 14-step task confidently but had fabricated three of the steps, with standard logs showing no errors. After 12 weeks of iteration, the engineer developed a four-field schema requiring each step log to capture expected outcome, observed outcome, a verification check, and an uncertainty signal. These fields directly address three recurring failure modes: silent HTTP failures masked by a parsed response, fabricated successes from partial results, and gradual confidence drift across multi-step tasks. The schema, shared publicly, replaces truncated and status-only logging with full-fidelity, severity-derived records that can surface fabrication before it reaches production.