Edib Imamovic

Response-quality grading on its own never catches the interesting failures. Action-sequence validation against an expected workflow, invariants on which tools get called for a given intent, custom policies beyond simple output checks; that's where the real agent bugs live. Getting that into the harness as a proper API rather than a checkbox was the thing we kept pushing for on the QA side.

Before i try it, somehow I have i feeling that it would be too much of a distruction? or im too old 🫣