A well written playbook on building practical evals for agents. Evals are the lens through which the AI community measures progress. The better the evals, the faster frontier models improve. Anything that can be measured can be improved, and recent model releases are clear
