[{"data":1,"prerenderedAt":340},["ShallowReactive",2],{"blog-post-/blogs/agent-self-evolution":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"image":11,"alt":12,"ogImage":11,"tags":13,"published":18,"body":19,"_type":333,"_id":334,"_source":335,"_file":336,"_stem":337,"_extension":338,"sitemap":339},"/blogs/agent-self-evolution","blogs",false,"","Agent自进化为什么是个伪命题","从OpenClaw团队访谈出发，分析为什么Agent框架目前不敢做真正的自进化——不是技术做不到，而是一旦认真做，就得放弃'稳定可控'这个核心卖点。","2026-05-18","/blogs-img/agent-self-evolution.jpg","AI Agent自进化主题手绘插画",[14,15,16,17],"AI","Agent","自进化","思考",true,{"type":20,"children":21,"toc":330},"root",[22,30,35,40,45,50,55,60,65,70,75,80,85,90,95,100,105,110,115,120,125,130,135,140,145,150,155,160,165,170,175,180,185,190,195,200,205,210,215,220,225,230,235,240,245,250,255,260,265,270,275,280,285,290,295,300,305,310,315,320,325],{"type":23,"tag":24,"props":25,"children":26},"element","p",{},[27],{"type":28,"value":29},"text","最近看了差评对 OpenClaw 团队 Josh 和 Michael 的专访，有一个问题特别有意思。",{"type":23,"tag":24,"props":31,"children":32},{},[33],{"type":28,"value":34},"记者问：\"你们为什么不加个自进化功能？让 Agent 能自己学习、改进？\"",{"type":23,"tag":24,"props":36,"children":37},{},[38],{"type":28,"value":39},"Josh 的回答是：\"开源生态的好处就是大家各自探索，官方不想替社区做选择。\"",{"type":23,"tag":24,"props":41,"children":42},{},[43],{"type":28,"value":44},"Michael 的回答是：\"现在太早了，很多东西还在变，今天 work 的东西明天模型一更新就不 work 了。\"",{"type":23,"tag":24,"props":46,"children":47},{},[48],{"type":28,"value":49},"我反复读这两段回答，觉得特别有意思。",{"type":23,"tag":24,"props":51,"children":52},{},[53],{"type":28,"value":54},"有意思的不是他们在说什么，而是他们没有说什么。",{"type":23,"tag":24,"props":56,"children":57},{},[58],{"type":28,"value":59},"我干智能体开发这一年多，见过太多类似的回答了。问 Agent 为什么不做自进化，行业里的人都在绕。绕的方式不太一样，但本质上都在回避同一个问题：",{"type":23,"tag":24,"props":61,"children":62},{},[63],{"type":28,"value":64},"不是不能，是不敢。",{"type":23,"tag":24,"props":66,"children":67},{},[68],{"type":28,"value":69},"先说为什么\"不能\"。",{"type":23,"tag":24,"props":71,"children":72},{},[73],{"type":28,"value":74},"我们今天搭的 Agent 框架，不管用什么技术栈，基本都是这个结构：",{"type":23,"tag":24,"props":76,"children":77},{},[78],{"type":28,"value":79},"工具调用写死、记忆检索用 Top-K 召回、任务编排走 DAG、输出格式靠 JSON Schema、Prompt 模板是字符串拼接。",{"type":23,"tag":24,"props":81,"children":82},{},[83],{"type":28,"value":84},"这套东西存在的意义是什么？",{"type":23,"tag":24,"props":86,"children":87},{},[88],{"type":28,"value":89},"让 AI 不要乱来。",{"type":23,"tag":24,"props":91,"children":92},{},[93],{"type":28,"value":94},"你花三个月搭的规则系统，每一行代码都在做同一件事：限制 AI 的自由度，把它按在你设计好的路径上跑。",{"type":23,"tag":24,"props":96,"children":97},{},[98],{"type":28,"value":99},"这不是 Agent，这是约束器。",{"type":23,"tag":24,"props":101,"children":102},{},[103],{"type":28,"value":104},"而自进化要求什么？要求你把自由度重新打开一点。让 Agent 自己决定用什么工具、怎么拆任务、记忆权重怎么调。",{"type":23,"tag":24,"props":106,"children":107},{},[108],{"type":28,"value":109},"但你一打开自由度，你的规则系统就废了。",{"type":23,"tag":24,"props":111,"children":112},{},[113],{"type":28,"value":114},"你一打开，AI 就有可能做错误决策。错了才能学，对吧？但错了就会出问题，你就得回头补规则。",{"type":23,"tag":24,"props":116,"children":117},{},[118],{"type":28,"value":119},"最后你会发现，你在做一件很矛盾的事：同时要求一个系统既听话又能自己改规矩。",{"type":23,"tag":24,"props":121,"children":122},{},[123],{"type":28,"value":124},"再说\"不敢\"的部分。",{"type":23,"tag":24,"props":126,"children":127},{},[128],{"type":28,"value":129},"所谓\"自进化\"现在在做什么？",{"type":23,"tag":24,"props":131,"children":132},{},[133],{"type":28,"value":134},"大多数产品演示的\"自进化\"，是这个流程：",{"type":23,"tag":24,"props":136,"children":137},{},[138],{"type":28,"value":139},"Agent 执行任务 → 拿到执行记录 → 塞回上下文 → 修改 Prompt → 调整 Memory",{"type":23,"tag":24,"props":141,"children":142},{},[143],{"type":28,"value":144},"然后它就\"学会\"了。",{"type":23,"tag":24,"props":146,"children":147},{},[148],{"type":28,"value":149},"但你仔细想想，这改变了什么？",{"type":23,"tag":24,"props":151,"children":152},{},[153],{"type":28,"value":154},"底层逻辑没变。你的工具调用还是写死的。你的 Top-K 召回还是那个算法。你的 DAG 还是那个拓扑结构。",{"type":23,"tag":24,"props":156,"children":157},{},[158],{"type":28,"value":159},"大模型会演，但 Agent 框架不会。",{"type":23,"tag":24,"props":161,"children":162},{},[163],{"type":28,"value":164},"这不叫自进化，这叫高阶版的 Prompt Engineering。你只是在上下文里塞了更多历史记录，让大模型表演\"我好像学到东西了\"。",{"type":23,"tag":24,"props":166,"children":167},{},[168],{"type":28,"value":169},"真正的自进化是什么？是 Agent 框架层能基于历史数据，优化自己的决策策略。",{"type":23,"tag":24,"props":171,"children":172},{},[173],{"type":28,"value":174},"但现在 Agent 框架的参数空间是什么？",{"type":23,"tag":24,"props":176,"children":177},{},[178],{"type":28,"value":179},"工具调用 → if-else。记忆管理 → 余弦相似度。任务编排 → 写死的 DAG。Prompt 模板 → 字符串拼接。",{"type":23,"tag":24,"props":181,"children":182},{},[183],{"type":28,"value":184},"这些东西是离散的、不可微的。你没法算梯度，没法做优化。强化学习想进来？先解决 Credit Assignment 再说——一个任务十几步工具调用，每步对最终结果贡献多少？你说得清楚吗？",{"type":23,"tag":24,"props":186,"children":187},{},[188],{"type":28,"value":189},"说不清楚。Reward Signal 不知道从哪来。",{"type":23,"tag":24,"props":191,"children":192},{},[193],{"type":28,"value":194},"所以最后大家都选择了一条路：假装在进化。",{"type":23,"tag":24,"props":196,"children":197},{},[198],{"type":28,"value":199},"那为什么行业不直说呢？",{"type":23,"tag":24,"props":201,"children":202},{},[203],{"type":28,"value":204},"我之前看到一句话，说的是段永平的不为清单。",{"type":23,"tag":24,"props":206,"children":207},{},[208],{"type":28,"value":209},"他说，他成功的秘诀不是知道什么该做，而是知道什么不该做。游戏不做，感冒药不做，多元化不做。",{"type":23,"tag":24,"props":211,"children":212},{},[213],{"type":28,"value":214},"知道不做什么，比知道做什么更重要。",{"type":23,"tag":24,"props":216,"children":217},{},[218],{"type":28,"value":219},"Agent 框架也一样。",{"type":23,"tag":24,"props":221,"children":222},{},[223],{"type":28,"value":224},"不做自进化，不是因为技术上做不到。而是因为一旦认真做自进化，你就得放弃 Agent 框架最核心的卖点——稳定可控。",{"type":23,"tag":24,"props":226,"children":227},{},[228],{"type":28,"value":229},"你的用户买的是什么？是\"我搭好规则，它按规矩跑，出问题我能修\"。不是\"它可能会变好，也可能变坏，反正我也不知道它怎么想的\"。",{"type":23,"tag":24,"props":231,"children":232},{},[233],{"type":28,"value":234},"这两件事在体验上是完全不同的。前者是工程产品，后者是实验品。",{"type":23,"tag":24,"props":236,"children":237},{},[238],{"type":28,"value":239},"你见过哪个工程产品宣传自己是\"可能变好也可能变坏的\"？",{"type":23,"tag":24,"props":241,"children":242},{},[243],{"type":28,"value":244},"Karpathy 前阵子说了一句话，我觉得特别准：",{"type":23,"tag":24,"props":246,"children":247},{},[248],{"type":28,"value":249},"\"2025 年或许是 Agent 的元年，但接下来的十年都将是 Agent 时代。\"",{"type":23,"tag":24,"props":251,"children":252},{},[253],{"type":28,"value":254},"什么意思？意思是现在 Agent 还不行，还得很长时间才能真正\"雇来干活\"。",{"type":23,"tag":24,"props":256,"children":257},{},[258],{"type":28,"value":259},"但我后来想想，这句话反过来也成立：",{"type":23,"tag":24,"props":261,"children":262},{},[263],{"type":28,"value":264},"正因为还需要十年，才说明这个方向是对的。",{"type":23,"tag":24,"props":266,"children":267},{},[268],{"type":28,"value":269},"正因为现在的 Agent 还在\"约束器\"阶段，才说明真正的\"自进化 Agent\"还在前面等着。",{"type":23,"tag":24,"props":271,"children":272},{},[273],{"type":28,"value":274},"我不是在否定 Agent 的价值。",{"type":23,"tag":24,"props":276,"children":277},{},[278],{"type":28,"value":279},"我是在说，承认当前的局限性，比假装突破更有价值。",{"type":23,"tag":24,"props":281,"children":282},{},[283],{"type":28,"value":284},"OpenClaw 为什么能火？不是因为它有多智能，而是因为它老老实实地承认：我是一个工具，一个可以控制的工具。用户愿意用，恰恰是因为它不会乱来。",{"type":23,"tag":24,"props":286,"children":287},{},[288],{"type":28,"value":289},"那些天天喊自进化、天天宣传 Agent 会自己学习的产品，你去看看他们的用户留存——估计大部分都是来薅羊毛的，真正愿意把关键任务交给它的，没几个。",{"type":23,"tag":24,"props":291,"children":292},{},[293],{"type":28,"value":294},"信任比智能更难建立。",{"type":23,"tag":24,"props":296,"children":297},{},[298],{"type":28,"value":299},"回到开头的问题：OpenClaw 为什么不做自进化？",{"type":23,"tag":24,"props":301,"children":302},{},[303],{"type":28,"value":304},"我现在觉得，Josh 和 Michael 的回答其实是诚实的。",{"type":23,"tag":24,"props":306,"children":307},{},[308],{"type":28,"value":309},"\"太早了\"，是对的。",{"type":23,"tag":24,"props":311,"children":312},{},[313],{"type":28,"value":314},"\"不想替社区做选择\"，也是对的。",{"type":23,"tag":24,"props":316,"children":317},{},[318],{"type":28,"value":319},"只是还有一个原因他们没说出来：",{"type":23,"tag":24,"props":321,"children":322},{},[323],{"type":28,"value":324},"他们比谁都清楚，一旦做了自进化，用户就不是在买工具了，而是在买不确定性。",{"type":23,"tag":24,"props":326,"children":327},{},[328],{"type":28,"value":329},"而这个世界上，愿意为不确定性付钱的人，永远比愿意为确定性付钱的人少。",{"title":7,"searchDepth":331,"depth":331,"links":332},2,[],"markdown","content:blogs:agent-self-evolution.md","content","blogs/agent-self-evolution.md","blogs/agent-self-evolution","md",{"loc":4},1779095310392]