问HN：有没有人通过自主工具实现递归自我改进？
感觉所有必要的组件终于可以用来构建一个自我强化的发展循环了。理论上，我们现在可以让像Claude Code或OpenClaw这样的工具监控一个git仓库，分析已完成工作的抽象，然后自主生成能够处理未来类似任务的新代理或技能。
有没有人成功运行过这样的循环？我很好奇这里是否有人将大部分时间从编写代码转向构建这些系统——本质上是在自举代理，使其能够从仓库历史中学习，以构建更好的代理。我很想听听那些在这方面不断突破的人。
查看原文
It feels like all the necessary components are finally available to build a self-reinforcing development loop.<p>Theoretically, we can now task tools like Claude Code or OpenClaw to monitor a git repo, analyze the abstractions in completed work, and then autonomously generate new agents or skills capable of handling similar tasks in the future.<p>Is anyone successfully running a loop like this? I’m curious if anyone here has shifted the majority of their time from writing code to crafting these systems—essentially bootstrapping agents that learn from the repo history to build better agents. I'd love to hear from those pushing the boundaries on this.