Why AI agents pass evals but fail in production

Outcome metrics miss step-level failures that compound across runs, creating a 35-point gap between benchmark scores and production success rates.