作者分享了用独立LLM强化/goal的实用技巧，用GPT-5.6-Sol做裁判，让GPT-5.6-Terra/Luna当子智能体，跑得更久更快。
推文作者认为OpenAI的/goal功能被低估，但需清晰目标。他利用单独的LLM（如Fable 5或GPT-5.6 Sol）构建了自己的/goal特性以减少LLM的偏见和奖励破解问题。GPT-5.6-Sol具备长时任务原生能力，无需/goal即可持续运行数天。作者将GPT-5.6-Terra/Luna作为子智能体，GPT-5.6-Sol作为顾问/裁判，使Codex订阅能更高效地完成复杂任务。
I actually think /goal is underrated. Yes, it does help to keep the agent going for longer, but it's...
I actually think /goal is underrated. Yes, it does help to keep the agent going for longer, but it's useless without a clearly set goal/criteria. Not to mention that LLMs (even frontier ones) have their own biases and reward hacking issues, which often get in the way. This is the reason I've built my own /goal feature that leverages a separate LLM to reduce those issues. I prefer using the best models like Fable 5 or GPT-5.6 Sol for this. With GPT-5.6-Sol having native capabilities to do longer-horizon tasks (as tobi points out) without the "itch" to pause every minute, /goal feels more like /utltragoal. I wouldn't give it up just yet, as it serves as an extra enforcing layer. GPT-5.6-Sol can run for days now. One powerful way I've been using my /goal feature is to leverage both GPT-5.6-Terra/Luna (heck, even GPT-5.5 works just as well) as subagents and GPT-5.6-Sol as the advisor/judge. I can squeeze more out of my Codex subscription when I do this, and it can do tasks a lot faster. tobi lutke @tobi OpenAI shipped /goal not too long ago. I feel like GPT-5.6-sol is the first model that just doesn't need it anymore? It just keeps on going until the thing is done. Really impressive. 🔗 View Quoted Tweet 💬 8 🔄 2 ❤️ 17 👀 2823 📊 8 ⚡