Testing an agent requires more than checking one answer. It means testing judgment across a changing sequence of decisions.