大语言模型天生不靠谱。上一篇文章聊了为什么需要 Agent 编排。这篇聊编排的核心：怎么约束 Agent 的行为。 跑了大半年 Agent 系统，我最大的体会：prompt 靠不住

I Only Do Two Things Now Record thoughts. Review drafts. And finally click publish. Recording thoughts means speaking into my phone. Reviewing means reading finished articles in Feishu documents and leaving comments paragraph by paragraph. Clicking publish is the very last step. Everything else — topic selection, writing, multi-round review, distribution — is handled by a team of AI agents. I call this system the “Content Collaboration Pipeline.” It’s not fully automatic.

现在写文章，我只做两件事 录随想，审稿。最后确认发布。 录随想是录一段语音。审稿是在飞书文档里看完成型的文章，逐段评论。确认发布是最后点一下。除