reference
GPT-6.1 agent
learned policy