Half of agent evaluation needs no LLM judge — and it's the half that catches the failures that actually hurt