智能体
x.com ·
@omarsar0
Nous Research 用 1393 个子代理重构百万行代码的观察
评论者关注 Nous Research 用 Hermes Agent 清理百万行 Python 代码的案例：1,393 个子代理运行 19 小时，代码库缩小 34.4%，据称节省近 200 万美元工程工时。作者认为 self-evolving skills 与 compounding engineering 是 harness 工程的重要技术，但指出同一方法未必能迁移到其他 harness，并质疑用 1.4K 子代理是否成本最优。
AI推荐理由 该帖转述 Nous Research 用 Hermes Agent 清理百万行 Python 代码的案例：1,393 个子代理、19 小时、代码库缩小 34.4%、节省近 200 万美元工程工时，并讨论 self-evolving skills、compounding engineering 以及 harness 差异导致方法难以迁移。案例具体、有量化数字，对 Agent harness 工程有参考价值。但数据来自被引用的官方博客，帖子本身是评论性转述，且作者对 1.4K 子代理是否必要提出疑问，效果未经独立验证。
74
浏览 3,167
点赞 17
评论 6
转发 2
正文
Recommended reading. This is probably one of the bigger subagent runs showing good results.
Two things caught my attention: self-evolving skills and compounding engineering.
Both are important techniques in harness engineering.
Overall, subagents reduced LOC significantly and delivered insane cost savings.
Hermes Agent is a proper agent harness with self-improving mechanisms, but I am not sure this same approach would work in other harnesses.
What works in one harness doesn't always transfer easily to others. This is why I am excited about custom harnesses. So much optimization is left out in these general-purpose coding harnesses.
Do pay attention to these differences in your harness.
My other observation was the use of a whopping ~1.4K subagents. Yes, it reduced costs, and it probably helped with parallelizing things. But could this run have been done cheaper using fewer subagents?
This is something to explore if you are a harness engineer.
I feel like this piece also highlights the importance of the harness and the customizations needed to provide the right context to agents when they need it.
Pay attention to how your agent harness can use self-evolving skills (encoding lessons and experience) to scale agents effectively.
Nous Research@NousResearch
New blog post:
We had a million lines of Python to clean up. On September 2nd [@Teknium](https://x.com/Teknium) asked Hermes Agent to do it.
1,393 subagents and nineteen hours later, the codebase was 34.4% smaller, saving us nearly $2m in engineering hours.
[nousresearch.com/refactoring-he…](https://nousresearch.com/refactoring-hermes-with-1393-agents)
[🔗 View Quoted Tweet](https://x.com/NousResearch/status/2099984561451028913)