video1.0<iframe src="https://www.loom.com/embed/8f06427f3e2d41fe824c5c1464a5316a" frameborder="0" width="2940" height="2205" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>22052940Loomhttps://www.loom.com22052940https://cdn.loom.com/sessions/thumbnails/8f06427f3e2d41fe824c5c1464a5316a-b508faa199bfe97a.gif89.499This Loom introduces Blackbox, a flight recorder for AI coding agents to detect and explain errors like hallucinated functions, fabricated dependencies, and incorrect claims that tests passed. The system identifies when an agent claims its tests passed despite conflicting results and catches fabricated issue references and non-existent functions, providing contextual evidence instead of vague alerts. Blackbox connects agent actions, claims, and supporting evidence into an interactive graph so developers can trace the exact decision that caused the mistake. The detection pipeline uses FastAPI, deterministic checks with a local semantic verification model called Glia, risk scoring with persistent SQLite storage, and streams updates via server-sent events.