How to benchmark AI agents across model versions

Stable benchmarks require fixed inputs, defined success criteria, and consistent instrumentation. Varying only one factor at a time makes results comparable.