学一招让Agent目标更靠谱
推文作者分享了一个提升Agent自主运行/goal效果的技巧：从历史会话中挖掘表现良好的目标，将这些洞察打包成自动化技能，供/goal工具复用。该方法可以解决LLM的奖励黑客行为、快速完成任务偏好等异常行为。作者已在编排器应用中构建了/goal的UI界面，并建议将这套做法作为Agent工具。
I also never set /goal by myself. The agent is probably better equipped with its context to help y...
I also never set /goal by myself. The agent is probably better equipped with its context to help you set a strong goal for longer autonomous runs. Smart to have it as a tool for agents. Exactly how I have it built in my orchestrator app. I even built a little UI for /goal in my orchestrator. Here is something you can try if you want to get better goals that agents follow. Mine your agent sessions, collect goals that performed well, and package those insights/best practices as a skill using an automation. That skill can then be reused with the /goal tool to set even stronger and more reliable goals. Doesn't get more meta than this. I am thinking of doing a live session on this if folks are interested. This is a simple workflow with tons of value and ways to optimize the results of /goal. It turns out that some of the weird behavior of LLM (e.g.g, reward hacking, bias to finish quickly, and other weird shortcuts) creep up a lot when using /goal, so you want to be extremely careful of that. I wrote a little more about better ways to set /goal here: x.com/omarsar0/statu… Tibo @thsottiaux Codex can see and set its own /goal. Everything we build, we build also as a tool for the agent. This is a generalization of meta prompting, where you let the agent set its own task based on your intent. 🔗 View Quoted Tweet 💬 3 🔄 3 ❤️ 21 👀 2265 📊 6 ⚡