http://mayf3.github.io/tags/ai/Recent content in AI on My HomeHugo -- gohugo.iozh-cnThu, 05 Nov 2026 00:00:00 +0800http://mayf3.github.io/post/2026-11-05-personal-agent-self-evolution/Thu, 05 Nov 2026 00:00:00 +0800http://mayf3.github.io/post/2026-11-05-personal-agent-self-evolution/Personal Agent：把不确定性纳入架构的自进化方法 开头：Agent 的问题不是“像不像函数” 传统软件习惯把能力封装成函数：输入明确、逻辑可重复、输出http://mayf3.github.io/post/2026-11-04-personal-agent-self-evolution/Wed, 04 Nov 2026 00:00:00 +0800http://mayf3.github.io/post/2026-11-04-personal-agent-self-evolution/Personal Agent、自进化与确定性—不确定性架构 引言：问题不是“模型够不够聪明” Personal Agent 的难点，不只是让模型生成更好的答案，而是让一个具有不确定性的执http://mayf3.github.io/post/2026-08-15-agent-permission-isolation/Sat, 15 Aug 2026 10:00:00 +0800http://mayf3.github.io/post/2026-08-15-agent-permission-isolation/Agent 一直用我的身份干活，直到我花两天做了权限隔离 我的 Agent 一直在用我的身份干活。不是比喻，是字面意思。 前段时间我把工作流和论坛都搭好了，正觉得一切http://mayf3.github.io/post/2026-08-15-agent-management-mindset-shift/Sat, 15 Aug 2026 08:00:00 +0800http://mayf3.github.io/post/2026-08-15-agent-management-mindset-shift/从跟每个 Agent 聊天到工作流管理：我的 Agent 管理心法转变 系列第 4 篇。前情：鉴权体系设计（为什么）→ 鉴权根治（怎么做）→ 编排进化（从单 Agent 到多 Agent）http://mayf3.github.io/post/2026-08-15-deepseek-harness-collision/Sat, 15 Aug 2026 08:00:00 +0800http://mayf3.github.io/post/2026-08-15-deepseek-harness-collision/我的轮子，被开源了。 昨天 DeepSeek 发布了 Harness，核心理念一句话：一切皆插件。而我手上这个做了很久的 Agent Core，一个类 Linux 内核的东西，内核极简、http://mayf3.github.io/post/2026-08-13-agent-auth-hardening/Thu, 13 Aug 2026 08:00:00 +0800http://mayf3.github.io/post/2026-08-13-agent-auth-hardening/把密钥藏起来：Agent 鉴权根治实录 上篇《Agent 鉴权体系设计：协作的地基》讲了：Agent 协作的第一性问题是身份——每个 Agent 得知道&quohttp://mayf3.github.io/post/2026-08-12-agent-orchestration-evolution/Wed, 12 Aug 2026 10:00:00 +0800http://mayf3.github.io/post/2026-08-12-agent-orchestration-evolution/用 Agent 干活的人，大概率都撞过同一堵墙：同一个任务，今天跑得顺，明天就翻车。你以为是运气问题，于是去调提示词，调来调去，它还是时好时坏。 我最早的http://mayf3.github.io/post/2026-08-07-agent-auth-system-design/Fri, 07 Aug 2026 08:00:00 +0800http://mayf3.github.io/post/2026-08-07-agent-auth-system-design/数字员工也需要工牌 每天早上，78 个 Agent 同时上线。有的推进工作流节点，有的管理 OKR 进度，有的在论坛讨论问题，有的审稿，有的发布文档。它们不坐工位、http://mayf3.github.io/post/2026-07-24-how-75-ai-agents-collaborate-five-layer-evolution/Fri, 24 Jul 2026 08:00:00 +0800http://mayf3.github.io/post/2026-07-24-how-75-ai-agents-collaborate-five-layer-evolution/Someone asked me: “You’ve got 75 AI agents (as of mid-July, counting both active and experimental ones). How do they actually interact? When Agent A finishes something, how does Agent B pick it up?” Great question. Because I didn’t have a clue at the start either. My initial thinking was simple: give each agent a role and some Skills, and they’d figure out how to work together. Turns out that’s nowhere near enough — Agent A writes a file that Agent B can’t read; Agent B finishes a step but Agent C has no idea it’s done.http://mayf3.github.io/post/2026-07-20-75-agents-collaboration-five-layer-evolution/Mon, 20 Jul 2026 08:00:00 +0800http://mayf3.github.io/post/2026-07-20-75-agents-collaboration-five-layer-evolution/有人问我：“你养了 75 个 AI Agent（截至 7 月中旬，含活跃和实验性角色），它们之间到底是怎么互动的？A 做完的事情 B 怎么接手？&rdhttp://mayf3.github.io/post/2026-07-15-ai-native-apps-agent-grown/Wed, 15 Jul 2026 10:00:00 +0800http://mayf3.github.io/post/2026-07-15-ai-native-apps-agent-grown/我给 AI Agent 团队搭了一堆系统：待办平台、学习系统、知识库。大部分没人用。 Agent 用不上是有原因的——我定义的"需求"根本不是他们需http://mayf3.github.io/post/2026-07-12-%E6%8A%8A-ai-%E6%94%BE%E5%9C%A8%E8%81%8A%E5%A4%A9%E8%BD%AF%E4%BB%B6%E9%87%8C-%E5%92%8C%E6%94%BE%E5%9C%A8%E7%94%B5%E8%84%91%E4%B8%8A%E6%98%AF%E4%B8%A4%E7%A7%8D%E4%B8%9C%E8%A5%BF/Sun, 12 Jul 2026 10:00:00 +0800http://mayf3.github.io/post/2026-07-12-%E6%8A%8A-ai-%E6%94%BE%E5%9C%A8%E8%81%8A%E5%A4%A9%E8%BD%AF%E4%BB%B6%E9%87%8C-%E5%92%8C%E6%94%BE%E5%9C%A8%E7%94%B5%E8%84%91%E4%B8%8A%E6%98%AF%E4%B8%A4%E7%A7%8D%E4%B8%9C%E8%A5%BF/在手机上和电脑上的 Agent 聊天之后，我就不用去电脑前了 我每天早上和晚上都用小爱同学播英文歌，给小孩做英语熏听。孩子不懂这门语言，但多听总归有好处—http://mayf3.github.io/post/design-review-format-upgrade/Fri, 03 Jul 2026 10:00:00 +0800http://mayf3.github.io/post/design-review-format-upgrade/一开始用 Copilot 时，Markdown 就是默认格式。后来 Agent 也一直往 Markdown 文件里写东西给我看——方案、分析、摘要，全是 Markdown。我也没多想，开发者http://mayf3.github.io/post/2026-05-25-agent-constraints-three-layer-system/Mon, 25 May 2026 08:00:00 +0800http://mayf3.github.io/post/2026-05-25-agent-constraints-three-layer-system/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http://mayf3.github.io/post/2026-05-17-i-wrote-code-for-17-years-then-ai-came/Sun, 17 May 2026 10:00:00 +0800http://mayf3.github.io/post/2026-05-17-i-wrote-code-for-17-years-then-ai-came/I’ve been writing code for 17 years, and recently, for the first time, I felt with real clarity that I might not make writing code my core competitive advantage anymore. AI has been changing too fast over the past few years. At first I treated it as a smarter search box. Then I found it could write code, analyze problems, call tools, even complete entire tasks on its own. For me, this wasn’t just exciting tech news — it changed how I work every single day.http://mayf3.github.io/post/2026-05-17-%E6%88%91%E5%86%99%E4%BA%8617%E5%B9%B4%E4%BB%A3%E7%A0%81%E7%84%B6%E5%90%8Eai%E6%9D%A5%E4%BA%86/Sun, 17 May 2026 10:00:00 +0800http://mayf3.github.io/post/2026-05-17-%E6%88%91%E5%86%99%E4%BA%8617%E5%B9%B4%E4%BB%A3%E7%A0%81%E7%84%B6%E5%90%8Eai%E6%9D%A5%E4%BA%86/写了 17 年代码，最近我第一次很清楚地感觉到，自己以后可能不会再以写代码为核心竞争力了。 过去几年 AI 的变化太快了。刚开始我还把它当成一个更聪明的搜http://mayf3.github.io/post/2026-05-07-how-my-shopping-assistant-grew-en/Thu, 07 May 2026 00:00:00 +0000http://mayf3.github.io/post/2026-05-07-how-my-shopping-assistant-grew-en/After I started using OpenClaw, I kept tinkering with different ways to use it—except when work got busy or I was too tired to bother. Recently I’ve been thinking about sharing some of my hands-on experiences from time to time. This time it’s about a shopping assistant that started as a simple shopping list and gradually grew into something that could search Xiaohongshu, compare prices, and add items to my cart.http://mayf3.github.io/post/2026-05-04-my-shopping-assistant/Mon, 04 May 2026 00:00:00 +0000http://mayf3.github.io/post/2026-05-04-my-shopping-assistant/开始用 OpenClaw 之后，我一直在折腾各种玩法，除了有时候工作比较忙或太累就顾不上。最近想时不时分享一些自己折腾的经验。 这次折腾的内容是：一个购物助手，http://mayf3.github.io/post/2026-04-26-teach-agent-gemini-avatar/Sun, 26 Apr 2026 22:42:00 +0800http://mayf3.github.io/post/2026-04-26-teach-agent-gemini-avatar/我从 OpenClaw 出来就一直在折腾各种玩法，除了有时候工作比较忙或太累就顾不上折腾。最近想时不时分享一些自己折腾的经验。 这次折腾的内容是：让图像生成专家