Evaluating agents beyond the first prompt

Single-turn scores overstate reliability—regressions, not missing features, are the real bottleneck.