An engineer built a 'Flight Recorder' for their Agent Control Plane to solve AI debugging's black box problem. When agents fail, traditional logs offer no root cause. This tool captures every interception—tool intent, policy verdict, and violation reason—into a local SQLite database, transforming guesswork into queryable data.

Standard LLM observability tracks tokens and latency, but misses governance context. The Flight Recorder records atomic decisions: what the agent tried to do, whether the kernel allowed it, and which rule triggered a block. This creates an audit trail for 'digital employees,' making failures reproducible and debuggable.

Using SQLite for 'Scale by Subtraction' means zero latency and cost. Developers can run SQL queries to analyze patterns, like blocked refund attempts or policy spikes after prompt updates. The kernel v1.0 ships with this enabled, turning AI debugging from a vibe check into a solvable engineering problem.