Demos often select favorable tasks and ignore failed runs. A useful agent must be evaluated on its stability, recovery capabilities, and cost.