Evals

AI Agent Benchmarks: What They Measure and Where They Fail

Practical guide to AI agent benchmarks (SWE-bench, tau-bench, WebArena, OSWorld, GAIA): what they measure, where they fail, and how to build a private eval.