#agents

Three eval modes — output, trajectory, and side effect — and the practical pattern we use when there's no single right answer.