{ "updatedAt": "2026-07-15T07:59:24Z", "frontier": [ { "id": "daily-2026-07-15", "title": "dorakuai 日报 · 2026-07-15：手机端跑27B模型，Meta AI物理奥赛满分", "category": "daily", "categoryName": "dorakuai 日报", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "今天", "date": "2026-07-15", "likes": 0, "featured": true, "summary": "Bonsai推出可在手机运行的27B级模型。Meta AI在亚洲物理奥赛理论考试中获满分，验证成本仍高昂。", "items": [], "overview": "Bonsai推出可在手机运行的27B级模型。Meta AI在亚洲物理奥赛理论考试中获满分，验证成本仍高昂。", "coverImage": "/assets/covers/daily-2026-07-15.jpg" }, { "id": "daily-2026-07-14", "title": "dorakuai 日报 · 2026-07-14：Agent 工作流、开源模型与端侧推理", "category": "daily", "categoryName": "dorakuai 日报", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "2026-07-14", "date": "2026-07-14", "likes": 0, "featured": false, "summary": "今日聚焦：编码 Agent 开始连接办公应用；开源多模态模型继续降低本地部署门槛；端侧推理与安全护栏仍是产品落地的主线。本页由 OpenClaw 日报流水线自动维护。", "coverImage": "/assets/covers/daily-2026-07-14.jpg", "items": [], "overview": "今日聚焦：编码 Agent 开始连接办公应用；开源多模态模型继续降低本地部署门槛；端侧推理与安全护栏仍是产品落地的主线。本页由 OpenClaw 日报流水线自动维护。" }, { "id": "agentic-detection", "title": "Agentic Detection：用一句话描述，AI 就在图里精确圈出目标，还能进行空间推测", "category": "research", "categoryName": "研究结果", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "本周", "likes": 0, "summary": "让视觉模型理解自然语言中的目标与空间关系，并以直观的框选方式返回结果。适合观察多模态 Agent 如何把感知与推理串成可操作流程。", "coverImage": "/assets/covers/agentic-detection.jpg", "overview": "让视觉模型理解自然语言中的目标与空间关系，并以直观的框选方式返回结果。适合观察多模态 Agent 如何把感知与推理串成可操作流程。", "items": [] }, { "id": "anthropic-flat-org", "title": "Anthropic CEO Dario Amodei 只有一个直接下属", "category": "news", "categoryName": "AI 资讯", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "本周", "likes": 0, "summary": "从组织设计观察 AI 公司的管理结构：高度扁平化能加快决策，也考验信息同步、授权边界与人才密度。", "coverImage": "/assets/covers/anthropic-flat-org.jpg", "overview": "从组织设计观察 AI 公司的管理结构：高度扁平化能加快决策，也考验信息同步、授权边界与人才密度。", "items": [] }, { "id": "diffusion-gemma", "title": "Google 开源扩散架构模型 DiffusionGemma：一次可同时生成 256 个 tokens", "category": "release", "categoryName": "产品发布", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "本周", "likes": 0, "summary": "扩散式语言模型尝试并行生成多个 token，为推理速度和交互形态提供新设计空间。", "coverImage": "/assets/covers/diffusion-gemma.jpg", "overview": "扩散式语言模型尝试并行生成多个 token，为推理速度和交互形态提供新设计空间。", "items": [] }, { "id": "live-translate", "title": "Google 发布实时语音翻译模型：能在 70 多种语言之间做到边听边译", "category": "product", "categoryName": "产品动态", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "本周", "likes": 0, "summary": "实时语音链路把识别、翻译与合成放进同一条低延迟路径，让跨语言沟通更接近自然对话。", "coverImage": "/assets/covers/live-translate.jpg", "overview": "实时语音链路把识别、翻译与合成放进同一条低延迟路径，让跨语言沟通更接近自然对话。", "items": [] }, { "id": "claude-guardrails", "title": "史上最强的 Claude 来了，Anthropic 却给它上了把锁", "category": "news", "categoryName": "AI 资讯", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "本周", "likes": 0, "summary": "模型能力越强，产品就越需要重新审视权限边界、安全护栏，以及用户对自动化的控制权。", "coverImage": "/assets/covers/claude-guardrails.jpg", "overview": "模型能力越强，产品就越需要重新审视权限边界、安全护栏，以及用户对自动化的控制权。", "items": [] }, { "id": "siri-ondevice", "title": "揭秘苹果全新 Siri AI 背后模型：如何将 200 亿参数的模型塞进手机里", "category": "research", "categoryName": "研究结果", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "近期", "likes": 0, "summary": "端侧模型需要在能力、内存、能耗与响应速度之间做精细取舍，模型压缩与异构计算因此成为关键。", "coverImage": "/assets/covers/siri-ondevice.jpg", "overview": "端侧模型需要在能力、内存、能耗与响应速度之间做精细取舍，模型压缩与异构计算因此成为关键。", "items": [] }, { "id": "notebooklm", "title": "NotebookLM 大升级：每个笔记本都配了能跑代码的云端电脑", "category": "product", "categoryName": "产品动态", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "近期", "likes": 1, "summary": "研究型工作区开始具备代码执行与多格式输出，把资料整理、分析计算和交付串成一条链路。", "coverImage": "/assets/covers/notebooklm.jpg", "overview": "研究型工作区开始具备代码执行与多格式输出，把资料整理、分析计算和交付串成一条链路。", "items": [] }, { "id": "ideogram-open", "title": "Ideogram 发布首个开源 AI 图像模型：文字渲染和版面控制拉到新高度", "category": "release", "categoryName": "产品发布", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "近期", "likes": 0, "summary": "开源图像模型加强文字渲染与版面控制，为设计工作流提供更稳定的基础能力。", "coverImage": "/assets/covers/ideogram-open.jpg", "overview": "开源图像模型加强文字渲染与版面控制，为设计工作流提供更稳定的基础能力。", "items": [] }, { "id": "gemma4", "title": "Google 发布 Gemma 4 12B 开源模型：16GB 笔记本跑全模态 AI", "category": "release", "categoryName": "产品发布", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "近期", "likes": 0, "summary": "更紧凑的全模态模型继续降低本地部署门槛，也为隐私优先和离线场景带来更多选择。", "coverImage": "/assets/covers/gemma4.jpg", "overview": "更紧凑的全模态模型继续降低本地部署门槛，也为隐私优先和离线场景带来更多选择。", "items": [] }, { "id": "codex-apps", "title": "Codex 发布重大更新：不再只是编码，开始连接更多办公应用", "category": "daily", "categoryName": "dorakuai 日报", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "近期", "likes": 0, "summary": "当编码代理连接更多工具，软件开发与日常知识工作的边界开始模糊，任务交付也更趋自动化。", "featured": false, "coverImage": "/assets/covers/codex-apps.jpg", "items": [], "overview": "当编码代理连接更多工具，软件开发与日常知识工作的边界开始模糊，任务交付也更趋自动化。" }, { "id": "koji-tutor", "title": "Koji：一个拒绝直接给答案的 AI 家教，强调图形化引导", "category": "daily", "categoryName": "dorakuai 日报", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "近期", "likes": 0, "summary": "好的教学型 AI 不只给答案，而是通过提问、图示和分步反馈，引导学习者建立自己的理解。", "featured": false, "coverImage": "/assets/covers/koji-tutor.jpg", "items": [], "overview": "好的教学型 AI 不只给答案，而是通过提问、图示和分步反馈，引导学习者建立自己的理解。" }, { "id": "agent-wallet", "title": "AI Agent 开始拥有自己的钱包：自动付款但不暴露真实银行卡信息", "category": "daily", "categoryName": "dorakuai 日报", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "近期", "likes": 0, "summary": "面向代理的支付通过令牌化与限额策略，让自动化系统能执行交易，同时保留用户控制权。", "featured": false, "coverImage": "/assets/covers/agent-wallet.jpg", "items": [], "overview": "面向代理的支付通过令牌化与限额策略，让自动化系统能执行交易，同时保留用户控制权。" } ], "toolbox": [ { "id": "little-language", "title": "Google 推出免费语言学习工具 Little Language Lessons，实时生成对话练习", "category": "language", "categoryName": "语言学习", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "精选", "likes": 0, "summary": "按学习场景即时生成短对话与表达练习，重点把词汇放进真实语境。", "coverImage": "/assets/covers/little-language.jpg", "overview": "按学习场景即时生成短对话与表达练习，重点把词汇放进真实语境。", "items": [] }, { "id": "promptfill", "title": "PromptFill：提示词填空器，让写提示词像填空一样简单", "category": "productivity", "categoryName": "效率工具", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "精选", "likes": 2, "summary": "把复杂提示词拆成可填写的结构化模块，降低反复试写和遗漏关键约束的成本。", "coverImage": "/assets/covers/promptfill.jpg", "overview": "把复杂提示词拆成可填写的结构化模块，降低反复试写和遗漏关键约束的成本。", "items": [] }, { "id": "firecrawl", "title": "Firecrawl 推出 AI 数据爬虫 Agent：描述需求即可自动采集", "category": "open", "categoryName": "开源工具", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "精选", "likes": 1, "summary": "将网页发现、抓取、清洗和结构化串成一条自动化链路，适合研究和数据整理场景。", "coverImage": "/assets/covers/firecrawl.jpg", "overview": "将网页发现、抓取、清洗和结构化串成一条自动化链路，适合研究和数据整理场景。", "items": [] }, { "id": "ps-web", "title": "Photoshop Chrome 扩展：直接在浏览器中编辑图片", "category": "image", "categoryName": "图像工具", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "精选", "likes": 1, "summary": "把常用图像编辑能力带到网页环境，减少浏览器与桌面软件之间的切换摩擦。", "coverImage": "/assets/covers/ps-web.jpg", "overview": "把常用图像编辑能力带到网页环境，减少浏览器与桌面软件之间的切换摩擦。", "items": [] }, { "id": "huxe", "title": "Huxe：AI 驱动的个性化语音交互电台，内容即时生成", "category": "model", "categoryName": "模型应用", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "精选", "likes": 0, "summary": "依据兴趣动态组织资讯与播客式内容，并通过语音交互继续追问和调整节目方向。", "coverImage": "/assets/covers/huxe.jpg", "overview": "依据兴趣动态组织资讯与播客式内容，并通过语音交互继续追问和调整节目方向。", "items": [] }, { "id": "mixboard", "title": "Google Mixboard：从空白开始协助构思创意的 AI 智能画板", "category": "image", "categoryName": "图像工具", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "精选", "likes": 1, "summary": "在可视化画布中组合参考、文字和图像，让灵感收集逐步转化为明确的创意方向。", "coverImage": "/assets/covers/mixboard.jpg", "overview": "在可视化画布中组合参考、文字和图像，让灵感收集逐步转化为明确的创意方向。", "items": [] }, { "id": "web-capture", "title": "Web Capture：一键抓取网页任意元素并转换成 React 代码", "category": "open", "categoryName": "开源工具", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "精选", "likes": 1, "summary": "识别页面元素的布局和样式并生成可继续编辑的组件代码，用于原型复现与前端学习。", "coverImage": "/assets/covers/web-capture.jpg", "overview": "识别页面元素的布局和样式并生成可继续编辑的组件代码，用于原型复现与前端学习。", "items": [] }, { "id": "drfonts", "title": "DrFonts：集字体生成、编辑与色彩管理于一体的 AI 字体工具", "category": "image", "categoryName": "图像工具", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "精选", "likes": 1, "summary": "围绕字形生成和风格一致性提供编辑流程，让非专业用户也能快速探索字体方案。", "coverImage": "/assets/covers/drfonts.jpg", "overview": "围绕字形生成和风格一致性提供编辑流程，让非专业用户也能快速探索字体方案。", "items": [] }, { "id": "lumi", "title": "Lumi：在阅读论文原文的同时获得 AI 辅助", "category": "productivity", "categoryName": "效率工具", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "精选", "likes": 0, "summary": "阅读过程中提供概念解释、段落总结和相关背景，尽量保留原文上下文与阅读节奏。", "coverImage": "/assets/covers/lumi.jpg", "overview": "阅读过程中提供概念解释、段落总结和相关背景，尽量保留原文上下文与阅读节奏。", "items": [] }, { "id": "pair-guide", "title": "Google PAIR：《以人为本的 AI 设计指南》", "category": "model", "categoryName": "模型应用", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "精选", "likes": 0, "summary": "从用户需求、反馈机制、错误恢复和信任校准等角度整理 AI 产品设计方法。", "coverImage": "/assets/covers/pair-guide.jpg", "overview": "从用户需求、反馈机制、错误恢复和信任校准等角度整理 AI 产品设计方法。", "items": [] }, { "id": "immersive-lang", "title": "浸入式语言学习助手：浏览网页时自动学习外语单词", "category": "language", "categoryName": "语言学习", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "精选", "likes": 3, "summary": "将词汇提示嵌入日常阅读页面，用轻量、连续的方式积累真实语境中的表达。", "coverImage": "/assets/covers/immersive-lang.jpg", "overview": "将词汇提示嵌入日常阅读页面，用轻量、连续的方式积累真实语境中的表达。", "items": [] }, { "id": "prompt-pilot", "title": "PromptPilot：提示词生成、优化、测试与管理工具", "category": "productivity", "categoryName": "效率工具", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "精选", "likes": 1, "summary": "围绕提示词版本、评测与复用建立工作台，适合需要稳定迭代模型任务的团队。", "coverImage": "/assets/covers/prompt-pilot.jpg", "overview": "围绕提示词版本、评测与复用建立工作台，适合需要稳定迭代模型任务的团队。", "items": [] } ], "deep": [ { "id": "openclaw-rl-07-policy-serving", "title": "OpenClaw-RL 源码笔记（7）：Policy Serving 策略服务架构", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-07-14", "likes": 0, "featured": true, "summary": "本系列的目的是：借着对 OpenClaw-RL 源码的学习，来梳理强化学习的一些相关概念和思想。所以，会有一些基础知识、扩展和发散，OpenClaw-RL 只是一个切入点。而且，因为整篇系列是一个整体，所以有些概念的解读/学习会在不同的文章中出现，还请大家谅解。", "overview": "本系列的目的是：借着对 OpenClaw-RL 源码的学习，来梳理强化学习的一些相关概念和思想。所以，会有一些基础知识、扩展和发散，OpenClaw-RL 只是一个切入点。而且，因为整篇系列是一个整体，所以有些概念的解读/学习会在不同的文章中出现，还请大家谅解。", "url": "https://www.cnblogs.com/rossiXYZ/p/20241201", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20241201", "series": "openclaw-rl", "seriesPart": 7, "space": "deep", "coverImage": "/assets/covers/openclaw-rl-07-policy-serving.jpg", "items": [ { "title": "OpenClaw-RL 源码系列", "titleZh": "📚 OpenClaw-RL 源码系列", "summary": "共 7 篇 · 站内专栏", "category": "专栏", "seriesKey": "openclaw-rl" }, { "title": "0x02 详细功能梳理", "titleZh": "0x02 详细功能梳理", "category": "章节", "url": "" }, { "title": "0x03 实现对比", "titleZh": "0x03 实现对比", "category": "章节", "url": "" }, { "title": "0x05 Rollout Allocation", "titleZh": "0x05 Rollout Allocation", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20241201", "external": true } ], "bodyPath": "/data/bodies/openclaw-rl-07-policy-serving.json" }, { "id": "qam-divl-loop", "title": "QAM 与 DIVL：评价-执行闭环如何完美配合", "category": "paper", "categoryName": "AI 论文", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-07-13", "likes": 0, "featured": true, "summary": "LWD 的核心主张是\"部署时持续改进\"，而这一主张在算法层面能否落地，取决于两个关键能力： **有没有一双火眼金睛，能从好坏参半的部署数据中看清真正的价值方向？有没有一身好功夫，能把价值方向转化为丝滑流畅的动作改进？** DIVL 和 QAM 就是为这两个问题而生的。", "overview": "LWD 的核心主张是\"部署时持续改进\"，而这一主张在算法层面能否落地，取决于两个关键能力： **有没有一双火眼金睛，能从好坏参半的部署数据中看清真正的价值方向？有没有一身好功夫，能把价值方向转化为丝滑流畅的动作改进？** DIVL 和 QAM 就是为这两个问题而生的。", "url": "https://www.cnblogs.com/rossiXYZ/p/21092694", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "21092694", "series": "robot-offline-rl", "seriesPart": 0, "space": "deep", "coverImage": "/assets/covers/qam-divl-loop.jpg", "items": [ { "title": "机器人离线 RL 系列", "titleZh": "📚 机器人离线 RL 系列", "summary": "共 5 篇 · 站内专栏", "category": "专栏", "seriesKey": "robot-offline-rl" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 为什么需要两个算法？", "titleZh": "0x01 为什么需要两个算法？", "category": "章节", "url": "" }, { "title": "0x02 LWD 论文算法", "titleZh": "0x02 LWD 论文算法", "category": "章节", "url": "" }, { "title": "0x02 DIVL：分布式价值评价系统", "titleZh": "0x02 DIVL：分布式价值评价系统", "category": "章节", "url": "" }, { "title": "0x03 QAM：伴随匹配策略执行系统", "titleZh": "0x03 QAM：伴随匹配策略执行系统", "category": "章节", "url": "" }, { "title": "0x04 协同机制：价值到行动的闭环", "titleZh": "0x04 协同机制：价值到行动的闭环", "category": "章节", "url": "" }, { "title": "0x05 总结", "titleZh": "0x05 总结", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/21092694", "external": true } ], "bodyPath": "/data/bodies/qam-divl-loop.json" }, { "id": "qam-flow-policy", "title": "QAM：基于伴随匹配的 Q-learning 流策略优化", "category": "paper", "categoryName": "AI 论文", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-07-11", "likes": 0, "featured": true, "summary": "现代视觉-语言-动作（VLA）模型常采用 Flow Matching 或 Diffusion 架构生成动作。这类模型表达能力强，可以生成平滑、多峰、符合物理规律的连续动作序列。但问题在于： **动作并非单步直接输出，而是通过多步去噪/ODE 积分逐步迭代生成。**", "overview": "现代视觉-语言-动作（VLA）模型常采用 Flow Matching 或 Diffusion 架构生成动作。这类模型表达能力强，可以生成平滑、多峰、符合物理规律的连续动作序列。但问题在于： **动作并非单步直接输出，而是通过多步去噪/ODE 积分逐步迭代生成。**", "url": "https://www.cnblogs.com/rossiXYZ/p/21092602", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "21092602", "series": "robot-offline-rl", "seriesPart": 0, "space": "deep", "coverImage": "/assets/covers/qam-flow-policy.jpg", "items": [ { "title": "机器人离线 RL 系列", "titleZh": "📚 机器人离线 RL 系列", "summary": "共 5 篇 · 站内专栏", "category": "专栏", "seriesKey": "robot-offline-rl" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x02 核心思想：Adjoint Matching", "titleZh": "0x02 核心思想：Adjoint Matching", "category": "章节", "url": "" }, { "title": "0x03 网络架构", "titleZh": "0x03 网络架构", "category": "章节", "url": "" }, { "title": "0x04 QAM 作者实现", "titleZh": "0x04 QAM 作者实现", "category": "章节", "url": "" }, { "title": "0x05 QAM 与相关方法的对比", "titleZh": "0x05 QAM 与相关方法的对比", "category": "章节", "url": "" }, { "title": "0x06 总结", "titleZh": "0x06 总结", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/21092602", "external": true } ], "bodyPath": "/data/bodies/qam-flow-policy.json" }, { "id": "divl-distributional", "title": "DIVL：分布隐式价值学习", "category": "paper", "categoryName": "AI 论文", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-07-09", "likes": 0, "featured": true, "summary": "DIVL（Distributional Implicit Value Learning）是 LWD 框架中承担核心价值评估使命的引擎。如果说 IQL 解决了\"如何在已知动作中提取价值上限\"的问题，那么 DIVL 则在此基础上迈出了关键一步——它不再输出单一的标量分数，而是输出一个完整的概率分布。这看似简单的变化，在真机部署场景中意味着从\"…", "overview": "DIVL（Distributional Implicit Value Learning）是 LWD 框架中承担核心价值评估使命的引擎。如果说 IQL 解决了\"如何在已知动作中提取价值上限\"的问题，那么 DIVL 则在此基础上迈出了关键一步——它不再输出单一的标量分数，而是输出一个完整的概率分布。这看似简单的变化，在真机部署场景中意味着从\"盲人摸象\"到\"全息透视\"的跨越。", "url": "https://www.cnblogs.com/rossiXYZ/p/21092548", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "21092548", "series": "robot-offline-rl", "seriesPart": 0, "space": "deep", "coverImage": "/assets/covers/divl-distributional.jpg", "items": [ { "title": "机器人离线 RL 系列", "titleZh": "📚 机器人离线 RL 系列", "summary": "共 5 篇 · 站内专栏", "category": "专栏", "seriesKey": "robot-offline-rl" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 从 IQL 到 DIVL：为什么标量不够用？", "titleZh": "0x01 从 IQL 到 DIVL：为什么标量不够用？", "category": "章节", "url": "" }, { "title": "0x02 DIVL 的核心设计：两种数学手段的融合", "titleZh": "0x02 DIVL 的核心设计：两种数学手段的融合", "category": "章节", "url": "" }, { "title": "0x03 网络架构", "titleZh": "0x03 网络架构", "category": "章节", "url": "" }, { "title": "0x04 Adaptive", "titleZh": "0x04 Adaptive", "category": "章节", "url": "" }, { "title": "0x05 Reward 信号：DIVL 如何识别\"正确\"", "titleZh": "0x05 Reward 信号：DIVL 如何识别\"正确\"", "category": "章节", "url": "" }, { "title": "0x06 DIVL 与 IQL 的对比", "titleZh": "0x06 DIVL 与 IQL 的对比", "category": "章节", "url": "" }, { "title": "0x07 总结", "titleZh": "0x07 总结", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/21092548", "external": true } ], "bodyPath": "/data/bodies/divl-distributional.json" }, { "id": "iql-implicit-q", "title": "IQL：离线强化学习的隐式价值提取", "category": "paper", "categoryName": "AI 论文", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-07-07", "likes": 0, "featured": true, "summary": "要理解 DIVL，必须先理解 IQL。IQL 是离线强化学习中的里程碑式工作，其精妙之处在于\"只用数据集里已有的数据，却能推断出比数据集更好的策略\"，全程不需要对未见过的动作进行采样。SERL/LWD 的训练通常从加载人类演示数据开始——如何从离线数据平滑地切换到在线实时学习，正是 IQL 解决的问题。", "overview": "要理解 DIVL，必须先理解 IQL。IQL 是离线强化学习中的里程碑式工作，其精妙之处在于\"只用数据集里已有的数据，却能推断出比数据集更好的策略\"，全程不需要对未见过的动作进行采样。SERL/LWD 的训练通常从加载人类演示数据开始——如何从离线数据平滑地切换到在线实时学习，正是 IQL 解决的问题。", "url": "https://www.cnblogs.com/rossiXYZ/p/21092483", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "21092483", "series": "robot-offline-rl", "seriesPart": 0, "space": "deep", "coverImage": "/assets/covers/iql-implicit-q.jpg", "items": [ { "title": "机器人离线 RL 系列", "titleZh": "📚 机器人离线 RL 系列", "summary": "共 5 篇 · 站内专栏", "category": "专栏", "seriesKey": "robot-offline-rl" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x02 网络架构", "titleZh": "0x02 网络架构", "category": "章节", "url": "" }, { "title": "0x03 核心参数解析", "titleZh": "0x03 核心参数解析", "category": "章节", "url": "" }, { "title": "0x04 核心训练循环", "titleZh": "0x04 核心训练循环", "category": "章节", "url": "" }, { "title": "0x05 IQL 价值流动的双向性", "titleZh": "0x05 IQL 价值流动的双向性", "category": "章节", "url": "" }, { "title": "0x06 策略网络的两种实现", "titleZh": "0x06 策略网络的两种实现", "category": "章节", "url": "" }, { "title": "0x07 核心优势总结", "titleZh": "0x07 核心优势总结", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/21092483", "external": true } ], "bodyPath": "/data/bodies/iql-implicit-q.json" }, { "id": "lwd-learning-while-deploying", "title": "LWD：机器人持续进化的 Learning while Deploying 框架", "category": "paper", "categoryName": "AI 论文", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-07-06", "likes": 0, "featured": true, "summary": "如果我们只看近几年的 VLA 模型，很容易形成一种直觉：只要把足够多的机器人数据喂给一个足够大的多模态模型，机器人就能获得广泛的通用能力。大规模离线数据确实让模型学会了看图像、理解指令、输出动作，也让机器人具备了初始的\"开箱即用\"能力。", "overview": "如果我们只看近几年的 VLA 模型，很容易形成一种直觉：只要把足够多的机器人数据喂给一个足够大的多模态模型，机器人就能获得广泛的通用能力。大规模离线数据确实让模型学会了看图像、理解指令、输出动作，也让机器人具备了初始的\"开箱即用\"能力。", "url": "https://www.cnblogs.com/rossiXYZ/p/21092330", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "21092330", "series": "robot-offline-rl", "seriesPart": 0, "space": "deep", "coverImage": "/assets/covers/lwd-learning-while-deploying.jpg", "items": [ { "title": "机器人离线 RL 系列", "titleZh": "📚 机器人离线 RL 系列", "summary": "共 5 篇 · 站内专栏", "category": "专栏", "seriesKey": "robot-offline-rl" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 背景：为什么需要 LWD？", "titleZh": "0x01 背景：为什么需要 LWD？", "category": "章节", "url": "" }, { "title": "0x02 系统架构：四大支柱", "titleZh": "0x02 系统架构：四大支柱", "category": "章节", "url": "" }, { "title": "0x03 核心 & 发展脉络", "titleZh": "0x03 核心 & 发展脉络", "category": "章节", "url": "" }, { "title": "0x04 两阶段训练流水线", "titleZh": "0x04 两阶段训练流水线", "category": "章节", "url": "" }, { "title": "0x05 分布式 Actor-Learner 架构", "titleZh": "0x05 分布式 Actor-Learner 架构", "category": "章节", "url": "" }, { "title": "0x06 论文算法解读", "titleZh": "0x06 论文算法解读", "category": "章节", "url": "" }, { "title": "0x07 LWD vs RECAP：两种技术路线的对比", "titleZh": "0x07 LWD vs RECAP：两种技术路线的对比", "category": "章节", "url": "" }, { "title": "0x08 为什么 LWD 能胜任长程任务", "titleZh": "0x08 为什么 LWD 能胜任长程任务", "category": "章节", "url": "" }, { "title": "0x09 总结：范式转移", "titleZh": "0x09 总结：范式转移", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/21092330", "external": true } ], "bodyPath": "/data/bodies/lwd-learning-while-deploying.json" }, { "id": "hil-serl-engineering", "title": "HIL-SERL 工程篇：人类在环架构与物理设计", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-07-03", "likes": 0, "featured": true, "summary": "HIL-SERL 能在真实机械臂上跑通 RL，靠的不是某个算法突破，而是整套工程系统设计：异步 Actor-Learner 解耦、SpaceMouse 实时干预、混合动作空间、阻抗控制底层安全——每一层都是纸上算法到真实机器人之间的必要桥梁。", "overview": "HIL-SERL 能在真实机械臂上跑通 RL，靠的不是某个算法突破，而是整套工程系统设计：异步 Actor-Learner 解耦、SpaceMouse 实时干预、混合动作空间、阻抗控制底层安全——每一层都是纸上算法到真实机器人之间的必要桥梁。", "url": "https://www.cnblogs.com/rossiXYZ/p/20899240", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20899240", "series": "hil-serl", "seriesPart": 0, "space": "deep", "coverImage": "/assets/covers/hil-serl-engineering.jpg", "items": [ { "title": "HIL-SERL 系列", "titleZh": "📚 HIL-SERL 系列", "summary": "共 4 篇 · 站内专栏", "category": "专栏", "seriesKey": "hil-serl" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 系统架构总览", "titleZh": "0x01 系统架构总览", "category": "章节", "url": "" }, { "title": "0x02 物理硬件设计", "titleZh": "0x02 物理硬件设计", "category": "章节", "url": "" }, { "title": "0x03 Human-in-the-Loop 机制", "titleZh": "0x03 Human-in-the-Loop 机制", "category": "章节", "url": "" }, { "title": "0x04 SpaceMouse 干预工程细节", "titleZh": "0x04 SpaceMouse 干预工程细节", "category": "章节", "url": "" }, { "title": "0x05 数据处理策略", "titleZh": "0x05 数据处理策略", "category": "章节", "url": "" }, { "title": "0x06 环境恢复与安全机制", "titleZh": "0x06 环境恢复与安全机制", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20899240", "external": true } ], "bodyPath": "/data/bodies/hil-serl-engineering.json" }, { "id": "hil-serl-hg-dagger", "title": "HIL-SERL 算法篇：HG-DAgger 与 RLPD 训练双阶段", "category": "paper", "categoryName": "AI 论文", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-07-01", "likes": 0, "featured": true, "summary": "HIL-SERL 的训练流程不是一个算法一跑到底，而是精心设计的两阶段接力：先用 HG-DAgger 做冷启动，让机器人快速学会基础操作；再用 RLPD 做持续优化，让机器人超越人类表现。这两个阶段共享相同的干预机制，但学习目标和数据处理方式截然不同。", "overview": "HIL-SERL 的训练流程不是一个算法一跑到底，而是精心设计的两阶段接力：先用 HG-DAgger 做冷启动，让机器人快速学会基础操作；再用 RLPD 做持续优化，让机器人超越人类表现。这两个阶段共享相同的干预机制，但学习目标和数据处理方式截然不同。", "url": "https://www.cnblogs.com/rossiXYZ/p/20899213", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20899213", "series": "hil-serl", "seriesPart": 0, "space": "deep", "coverImage": "/assets/covers/hil-serl-hg-dagger.jpg", "items": [ { "title": "HIL-SERL 系列", "titleZh": "📚 HIL-SERL 系列", "summary": "共 4 篇 · 站内专栏", "category": "专栏", "seriesKey": "hil-serl" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 HG-DAgger 的核心思想：人类门控的数据聚合", "titleZh": "0x01 HG-DAgger 的核心思想：人类门控的数据聚合", "category": "章节", "url": "" }, { "title": "0x02 HG-DAgger 训练范式全解析", "titleZh": "0x02 HG-DAgger 训练范式全解析", "category": "章节", "url": "" }, { "title": "0x03 RLPD 训练范式全解析", "titleZh": "0x03 RLPD 训练范式全解析", "category": "章节", "url": "" }, { "title": "0x04 对比", "titleZh": "0x04 对比", "category": "章节", "url": "" }, { "title": "0x05 HIL-SERL 的总体支柱", "titleZh": "0x05 HIL-SERL 的总体支柱", "category": "章节", "url": "" }, { "title": "0x06 总结", "titleZh": "0x06 总结", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20899213", "external": true } ], "bodyPath": "/data/bodies/hil-serl-hg-dagger.json" }, { "id": "hil-serl-dqn-sac", "title": "HIL-SERL 算法篇：DQN + SAC 混合架构实现哲学", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-06-30", "likes": 0, "featured": true, "summary": "HIL-SERL 不是单一算法，而是一套高度集成的混合 RL 系统。它用 SAC 给机器人灵动的手臂、用 DQN 给机器人果断的夹爪、用二值分类器替代手工奖励设计、用人类干预数据弥合 RL 探索与 BC 模仿之间的鸿沟。这篇我们从算法角度逐层拆解它的设计逻辑。", "overview": "HIL-SERL 不是单一算法，而是一套高度集成的混合 RL 系统。它用 SAC 给机器人灵动的手臂、用 DQN 给机器人果断的夹爪、用二值分类器替代手工奖励设计、用人类干预数据弥合 RL 探索与 BC 模仿之间的鸿沟。这篇我们从算法角度逐层拆解它的设计逻辑。", "url": "https://www.cnblogs.com/rossiXYZ/p/20899192", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20899192", "series": "hil-serl", "seriesPart": 0, "space": "deep", "coverImage": "/assets/covers/hil-serl-dqn-sac.jpg", "items": [ { "title": "HIL-SERL 系列", "titleZh": "📚 HIL-SERL 系列", "summary": "共 4 篇 · 站内专栏", "category": "专栏", "seriesKey": "hil-serl" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 HIL-SERL 的总体算法思路", "titleZh": "0x01 HIL-SERL 的总体算法思路", "category": "章节", "url": "" }, { "title": "0x02 核心算法：RLPD + 人类干预", "titleZh": "0x02 核心算法：RLPD + 人类干预", "category": "章节", "url": "" }, { "title": "0x03 混合动作空间：SAC + DQN", "titleZh": "0x03 混合动作空间：SAC + DQN", "category": "章节", "url": "" }, { "title": "0x04 GraspCritic：夹爪离散决策网络", "titleZh": "0x04 GraspCritic：夹爪离散决策网络", "category": "章节", "url": "" }, { "title": "0x05 奖励系统：二值分类器", "titleZh": "0x05 奖励系统：二值分类器", "category": "章节", "url": "" }, { "title": "0x06 SACAgentHybridSingleArm：单臂混合动作 SAC Agent", "titleZh": "0x06 SACAgentHybridSingleArm：单臂混合动作 SAC Agent", "category": "章节", "url": "" }, { "title": "0x07 训练稳定性机制", "titleZh": "0x07 训练稳定性机制", "category": "章节", "url": "" }, { "title": "0x08 总结：算法的基因与局限", "titleZh": "0x08 总结：算法的基因与局限", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20899192", "external": true } ], "bodyPath": "/data/bodies/hil-serl-dqn-sac.json" }, { "id": "hil-serl-overview", "title": "HIL-SERL：人类在环驱动的具身智能进化框架", "category": "paper", "categoryName": "AI 论文", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-06-29", "likes": 0, "featured": true, "summary": "真实世界中的机器人学习，不应该让机器人孤军奋战。人类纠偏不是训练之外的异常——它是高价值训练数据的关键来源。HIL-SERL（Human-in-the-loop SERL）的核心主张是： **让机器人自己试，人类只在关键时刻扶一把；而这一下\"扶正\"，会被系统转化为策略改进信号。**", "overview": "真实世界中的机器人学习，不应该让机器人孤军奋战。人类纠偏不是训练之外的异常——它是高价值训练数据的关键来源。HIL-SERL（Human-in-the-loop SERL）的核心主张是： **让机器人自己试，人类只在关键时刻扶一把；而这一下\"扶正\"，会被系统转化为策略改进信号。**", "url": "https://www.cnblogs.com/rossiXYZ/p/20899158", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20899158", "series": "hil-serl", "seriesPart": 0, "space": "deep", "coverImage": "/assets/covers/hil-serl-overview.jpg", "items": [ { "title": "HIL-SERL 系列", "titleZh": "📚 HIL-SERL 系列", "summary": "共 4 篇 · 站内专栏", "category": "专栏", "seriesKey": "hil-serl" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 真机 RL 的\"最后一公里\"", "titleZh": "0x01 真机 RL 的\"最后一公里\"", "category": "章节", "url": "" }, { "title": "0x02 设计哲学：三句话讲清 HIL-SERL 怎么想", "titleZh": "0x02 设计哲学：三句话讲清 HIL-SERL 怎么想", "category": "章节", "url": "" }, { "title": "0x03 SERL vs HIL-SERL：演进了什么", "titleZh": "0x03 SERL vs HIL-SERL：演进了什么", "category": "章节", "url": "" }, { "title": "0x04 系统架构：两个进程 + 三层 Wrapper", "titleZh": "0x04 系统架构：两个进程 + 三层 Wrapper", "category": "章节", "url": "" }, { "title": "0x05 核心机制：每个组件解决什么问题", "titleZh": "0x05 核心机制：每个组件解决什么问题", "category": "章节", "url": "" }, { "title": "0x06 训练生命周期", "titleZh": "0x06 训练生命周期", "category": "章节", "url": "" }, { "title": "0x07 SERL vs HIL-SERL 系统级对比", "titleZh": "0x07 SERL vs HIL-SERL 系统级对比", "category": "章节", "url": "" }, { "title": "0x08 从 HIL-SERL 到 LWD：范式如何继续演进", "titleZh": "0x08 从 HIL-SERL 到 LWD：范式如何继续演进", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20899158", "external": true } ], "bodyPath": "/data/bodies/hil-serl-overview.json" }, { "id": "serl-05-engineering", "title": "【机器人 / 强化学习】SERL：让真机强化学习从“难用”走向“可复现”的强化学习框架 ----（5）工程篇", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-06-28", "likes": 0, "featured": false, "summary": "当我们谈论具身智能时，最容易忽略的是： **在真实机器人上跑 RL 本身就是一道工程难题**。策略网络可以在仿真中一日千里地进化，但一旦部署到物理世界，采样效率低、硬件易损、重置成本高、奖励设计难——每一个问题都可能让训练无法持续。SERL 正是为了解决这些问题而生的工程框架。", "overview": "当我们谈论具身智能时，最容易忽略的是： **在真实机器人上跑 RL 本身就是一道工程难题**。策略网络可以在仿真中一日千里地进化，但一旦部署到物理世界，采样效率低、硬件易损、重置成本高、奖励设计难——每一个问题都可能让训练无法持续。SERL 正是为了解决这些问题而生的工程框架。", "url": "https://www.cnblogs.com/rossiXYZ/p/20878136", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20878136", "series": "serl", "seriesPart": 5, "space": "deep", "coverImage": "/assets/covers/serl-05-engineering.jpg", "items": [ { "title": "SERL 真机 RL 系列", "titleZh": "📚 SERL 真机 RL 系列", "summary": "共 5 篇 · 站内专栏", "category": "专栏", "seriesKey": "serl" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 SERL 要解决什么核心问题？", "titleZh": "0x01 SERL 要解决什么核心问题？", "category": "章节", "url": "" }, { "title": "0x02 系统架构：三层解耦的通用适配器", "titleZh": "0x02 系统架构：三层解耦的通用适配器", "category": "章节", "url": "" }, { "title": "0x03 控制层：阻抗控制是 SERL 的物理底座", "titleZh": "0x03 控制层：阻抗控制是 SERL 的物理底座", "category": "章节", "url": "" }, { "title": "0x04 数据层：视觉记忆与采样优化", "titleZh": "0x04 数据层：视觉记忆与采样优化", "category": "章节", "url": "" }, { "title": "0x05 相对坐标系：让策略不再\"路痴\"", "titleZh": "0x05 相对坐标系：让策略不再\"路痴\"", "category": "章节", "url": "" }, { "title": "0x06 Reset-Free Training：让机器人自己\"重置考场\"", "titleZh": "0x06 Reset-Free Training：让机器人自己\"重置考场\"", "category": "章节", "url": "" }, { "title": "0x07 总结：SERL 的工程遗产", "titleZh": "0x07 总结：SERL 的工程遗产", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20878136", "external": true } ], "bodyPath": "/data/bodies/serl-05-engineering.json" }, { "id": "serl-04-drq-vice", "title": "【机器人 / 强化学习】SERL：让真机强化学习从“难用”走向“可复现”的强化学习框架 ----（4）算法篇（DrQ vs VICE）", "category": "paper", "categoryName": "AI 论文", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-06-26", "likes": 0, "featured": false, "summary": "某些任务可以用机器人状态直接定义，例如 PCB 插入可以根据末端或物体位置设计奖励；在这种情况下，奖励函数可以由一个二元分类器提供，该分类器接收状态观测 s 并输出一个二元” 事件”e发生的概率，对应于任务的成功完成。奖励随后由 r(s) = log p(e\\|s) 给出。", "overview": "· 机器人必须通过像素看世界，DrQ 解决了视觉特征提取的泛化难题。\n· VICE 解决了\"真实世界没有代码奖励\"的问题，让机器人拥有了\"成就感\"。", "url": "https://www.cnblogs.com/rossiXYZ/p/20850061", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20850061", "series": "serl", "seriesPart": 4, "space": "deep", "coverImage": "/assets/covers/serl-04-drq-vice.jpg", "items": [ { "title": "SERL 真机 RL 系列", "titleZh": "📚 SERL 真机 RL 系列", "summary": "共 5 篇 · 站内专栏", "category": "专栏", "seriesKey": "serl" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 奖励与自动化：从手写 reward 到 learned reward", "titleZh": "0x01 奖励与自动化：从手写 reward 到 learned reward", "category": "章节", "url": "" }, { "title": "0x02 DrQ vs VICE", "titleZh": "0x02 DrQ vs VICE", "category": "章节", "url": "" }, { "title": "0x03 DrQ：让视觉 RL 更稳健", "titleZh": "0x03 DrQ：让视觉 RL 更稳健", "category": "章节", "url": "" }, { "title": "0x04 VICE", "titleZh": "0x04 VICE", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20850061", "external": true } ], "bodyPath": "/data/bodies/serl-04-drq-vice.json" }, { "id": "serl-03-rlpd", "title": "【机器人 / 强化学习】SERL：让真机强化学习从“难用”走向“可复现”的强化学习框架 ----（3）算法篇（RLPD）", "category": "paper", "categoryName": "AI 论文", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-06-25", "likes": 0, "featured": false, "summary": "SERL 的核心使命是：在真实世界中，让机器人在 20-40 分钟内学会高精度的机械操作。它通过集成 SAC、RLPD、DrQ 和 VICE，将原本需要数百万次尝试的 RL，压缩到了人类演示水平的量级。", "overview": "SERL 的核心使命是：在真实世界中，让机器人在 20-40 分钟内学会高精度的机械操作。它通过集成 SAC、RLPD、DrQ 和 VICE，将原本需要数百万次尝试的 RL，压缩到了人类演示水平的量级。", "url": "https://www.cnblogs.com/rossiXYZ/p/20781196", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20781196", "series": "serl", "seriesPart": 3, "space": "deep", "coverImage": "/assets/covers/serl-03-rlpd.jpg", "items": [ { "title": "SERL 真机 RL 系列", "titleZh": "📚 SERL 真机 RL 系列", "summary": "共 5 篇 · 站内专栏", "category": "专栏", "seriesKey": "serl" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 基础 & 背景", "titleZh": "0x01 基础 & 背景", "category": "章节", "url": "" }, { "title": "0x02 RLPD 基础", "titleZh": "0x02 RLPD 基础", "category": "章节", "url": "" }, { "title": "0x03 BC", "titleZh": "0x03 BC", "category": "章节", "url": "" }, { "title": "0x04 High UTD & 稳定性机制", "titleZh": "0x04 High UTD & 稳定性机制", "category": "章节", "url": "" }, { "title": "0x05 Layer Normalization", "titleZh": "0x05 Layer Normalization", "category": "章节", "url": "" }, { "title": "0x06 Soft Update & REDQ", "titleZh": "0x06 Soft Update & REDQ", "category": "章节", "url": "" }, { "title": "0x07 SAC vs RLPD", "titleZh": "0x07 SAC vs RLPD", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20781196", "external": true } ], "bodyPath": "/data/bodies/serl-03-rlpd.json" }, { "id": "serl-02-sac", "title": "【机器人 / 强化学习】SERL：让真机强化学习从“难用”走向“可复现”的强化学习框架 ---- （2）算法篇（SAC）", "category": "paper", "categoryName": "AI 论文", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-06-23", "likes": 0, "featured": false, "summary": "SAC 的\"Soft\"之源：传统强化学习目标是最大化累积奖励 ∑γ∧trt，SAC 在这个目标函数中额外增加了一项。SAC 解决了连续空间下的动作控制问题，它的\"灵魂\" 就是在于熵（Entropy）的引入。", "overview": "· **本系列的最终目标是“通过一系列相关项目/算法的解读，来深入学习/分析/反推 LWD（Learning while Deploying）这篇论文的机理和可能实现”**。之所以从SERL入手，是因为 SERL，HIL-SERL，SOP（没有开源）都是罗剑岚博士的一系列论文，可以从中管窥作者的思路脉络。\n· 本文依然是从工程/论文进行反推，还请读者不吝指出问题，多谢。", "url": "https://www.cnblogs.com/rossiXYZ/p/20745451", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20745451", "series": "serl", "seriesPart": 2, "space": "deep", "coverImage": "/assets/covers/serl-02-sac.jpg", "items": [ { "title": "SERL 真机 RL 系列", "titleZh": "📚 SERL 真机 RL 系列", "summary": "共 5 篇 · 站内专栏", "category": "专栏", "seriesKey": "serl" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 基础 & 背景", "titleZh": "0x01 基础 & 背景", "category": "章节", "url": "" }, { "title": "0x02 演化脉络", "titleZh": "0x02 演化脉络", "category": "章节", "url": "" }, { "title": "0x03 网络结构", "titleZh": "0x03 网络结构", "category": "章节", "url": "" }, { "title": "0x04 实现", "titleZh": "0x04 实现", "category": "章节", "url": "" }, { "title": "0x05 特色功能", "titleZh": "0x05 特色功能", "category": "章节", "url": "" }, { "title": "0x06 SAC 的工作流程", "titleZh": "0x06 SAC 的工作流程", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20745451", "external": true } ], "bodyPath": "/data/bodies/serl-02-sac.json" }, { "id": "serl-01-overview", "title": "【机器人 / 强化学习】SERL：让真机强化学习从“难用”走向“可复现”的强化学习框架 ----（1）全景篇", "category": "paper", "categoryName": "AI 论文", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-06-22", "likes": 0, "featured": true, "summary": "SERL（Sample-Efficient Robotic Reinforcement Learning）为我们证明了一件非常关键的事情：只要算法、控制、奖励、重置、数据管理和分布式训练实现足够细致，真实机器人强化学习可以在几十分钟级别完成复杂技能学习。", "overview": "SERL（Sample-Efficient Robotic Reinforcement Learning）为我们证明了一件非常关键的事情：只要算法、控制、奖励、重置、数据管理和分布式训练实现足够细致，真实机器人强化学习可以在几十分钟级别完成复杂技能学习。", "url": "https://www.cnblogs.com/rossiXYZ/p/20711693", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20711693", "series": "serl", "seriesPart": 1, "space": "deep", "coverImage": "/assets/covers/serl-01-overview.jpg", "items": [ { "title": "SERL 真机 RL 系列", "titleZh": "📚 SERL 真机 RL 系列", "summary": "共 5 篇 · 站内专栏", "category": "专栏", "seriesKey": "serl" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 基础知识 & 背景", "titleZh": "0x01 基础知识 & 背景", "category": "章节", "url": "" }, { "title": "0x02 SERL 的总体架构：一个垂直整合的真机 RL 栈", "titleZh": "0x02 SERL 的总体架构：一个垂直整合的真机 RL 栈", "category": "章节", "url": "" }, { "title": "0x03 SERL 的算法脉络", "titleZh": "0x03 SERL 的算法脉络", "category": "章节", "url": "" }, { "title": "0x04 SERL 在机器人 RL 方法谱系中的多维定位", "titleZh": "0x04 SERL 在机器人 RL 方法谱系中的多维定位", "category": "章节", "url": "" }, { "title": "0x05 SERL 的工程脉络", "titleZh": "0x05 SERL 的工程脉络", "category": "章节", "url": "" }, { "title": "0x06 SERL 与 HIL-SERL、LWD 的演进关系", "titleZh": "0x06 SERL 与 HIL-SERL、LWD 的演进关系", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20711693", "external": true } ], "bodyPath": "/data/bodies/serl-01-overview.json" }, { "id": "openclaw-rl-06-rollout", "title": "【Agentic RL / 强化学习 / OPD】OpenClaw-RL 源码阅读笔记 --- (6)--- Rollout", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-06-18", "likes": 0, "featured": false, "summary": "本系列的目的是：借着对 OpenClaw-RL 源码的学习，来梳理强化学习的一些相关概念和思想。所以，会有一些基础知识、扩展和发散，OpenClaw-RL 只是一个切入点。而且，因为整篇系列是一个整体，所以有些概念的解读/学习会在不同的文章中出现，还请大家谅解。", "overview": "本系列的目的是：借着对 OpenClaw-RL 源码的学习，来梳理强化学习的一些相关概念和思想。所以，会有一些基础知识、扩展和发散，OpenClaw-RL 只是一个切入点。而且，因为整篇系列是一个整体，所以有些概念的解读/学习会在不同的文章中出现，还请大家谅解。", "url": "https://www.cnblogs.com/rossiXYZ/p/20241056", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20241056", "series": "openclaw-rl", "seriesPart": 6, "space": "deep", "coverImage": "/assets/covers/openclaw-rl-06-rollout.jpg", "items": [ { "title": "OpenClaw-RL 源码系列", "titleZh": "📚 OpenClaw-RL 源码系列", "summary": "共 7 篇 · 站内专栏", "category": "专栏", "seriesKey": "openclaw-rl" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 Rollout基础", "titleZh": "0x01 Rollout基础", "category": "章节", "url": "" }, { "title": "0x02 OpenClaw-RL Rollout基础", "titleZh": "0x02 OpenClaw-RL Rollout基础", "category": "章节", "url": "" }, { "title": "0x03 OpenClaw-RL Rollout 实现", "titleZh": "0x03 OpenClaw-RL Rollout 实现", "category": "章节", "url": "" }, { "title": "0x04 AsyncRolloutWorker", "titleZh": "0x04 AsyncRolloutWorker", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20241056", "external": true } ], "bodyPath": "/data/bodies/openclaw-rl-06-rollout.json" }, { "id": "miles-02-key-tech", "title": "【Agentic RL / 强化学习框架】Miles 项目技术分析---（2）--- 关键技术", "category": "paper", "categoryName": "AI 论文", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-06-16", "likes": 0, "featured": false, "summary": "在 Agentic RL 中，Agent 开发者需要处理的不只是\"如何调用工具、如何解析结果\"，还有一连串的训练基础设施问题，而agentic\\_tool\\_call 是将 Agent 业务逻辑与 RL 训练基础设施解耦的适配器。", "overview": "在 Agentic RL 中，Agent 开发者需要处理的不只是\"如何调用工具、如何解析结果\"，还有一连串的训练基础设施问题，而agentic\\_tool\\_call 是将 Agent 业务逻辑与 RL 训练基础设施解耦的适配器。", "url": "https://www.cnblogs.com/rossiXYZ/p/20524400", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20524400", "series": "miles", "seriesPart": 2, "space": "deep", "coverImage": "/assets/covers/miles-02-key-tech.jpg", "items": [ { "title": "Miles 框架系列", "titleZh": "📚 Miles 框架系列", "summary": "共 2 篇 · 站内专栏", "category": "专栏", "seriesKey": "miles" }, { "title": "0x04 训推一致性频谱：从全异步到比特级一致的逐层递进", "titleZh": "0x04 训推一致性频谱：从全异步到比特级一致的逐层递进", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20524400", "external": true } ], "bodyPath": "/data/bodies/miles-02-key-tech.json" }, { "id": "miles-01-overview", "title": "【Agentic RL / 强化学习框架】Miles 项目技术分析---（1）--- 总体", "category": "paper", "categoryName": "AI 论文", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-06-15", "likes": 0, "featured": true, "summary": "Miles的工作以B（新增）+C（改核心）为主，纯利用扩展点的工作只是冰山一角。真正的价值在于Slime完全没预见的“会话化正确性“和“训推致性治理“等新架构层，是对计算核心、传输层、正确性保障、可观测性做了全面的工程建设。---- 具体体现为：", "overview": "Miles的工作以B（新增）+C（改核心）为主，纯利用扩展点的工作只是冰山一角。真正的价值在于Slime完全没预见的“会话化正确性“和“训推致性治理“等新架构层，是对计算核心、传输层、正确性保障、可观测性做了全面的工程建设。---- 具体体现为：", "url": "https://www.cnblogs.com/rossiXYZ/p/20524102", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20524102", "series": "miles", "seriesPart": 1, "space": "deep", "coverImage": "/assets/covers/miles-01-overview.jpg", "items": [ { "title": "Miles 框架系列", "titleZh": "📚 Miles 框架系列", "summary": "共 2 篇 · 站内专栏", "category": "专栏", "seriesKey": "miles" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 基础", "titleZh": "0x01 基础", "category": "章节", "url": "" }, { "title": "0x02 从 Slime 说起", "titleZh": "0x02 从 Slime 说起", "category": "章节", "url": "" }, { "title": "0x03 Miles 的升级", "titleZh": "0x03 Miles 的升级", "category": "章节", "url": "" }, { "title": "0x04 Miles 的逻辑分层", "titleZh": "0x04 Miles 的逻辑分层", "category": "章节", "url": "" }, { "title": "0x05 Miles 每项工作解决的问题", "titleZh": "0x05 Miles 每项工作解决的问题", "category": "章节", "url": "" }, { "title": "0x06 对比", "titleZh": "0x06 对比", "category": "章节", "url": "" }, { "title": "0x07 Agentic RL 8 大关键维度", "titleZh": "0x07 Agentic RL 8 大关键维度", "category": "章节", "url": "" }, { "title": "0xEE 广告", "titleZh": "0xEE 广告", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20524102", "external": true } ], "bodyPath": "/data/bodies/miles-01-overview.json" }, { "id": "uni-agent-02-key-tech", "title": "【强化学习框架】Uni-Agent 深度技术分析（2）--- 关键技术", "category": "paper", "categoryName": "AI 论文", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-06-11", "likes": 0, "featured": false, "summary": "项目的核心主张可以概括为 **\"推理 = 训练，同一入口\"**——把\"Agent 如何与环境交互并产出 token-level 训练数据\"封装成一套可复用、可扩展、可训练的工程系统。无论你是跑 1000 个并行的 SWE-Bench 推理任务，还是进行 GRPO/GSPO 强化学习训练，底层走的是同一套 `AgentInteraction…", "overview": "项目的核心主张可以概括为 **\"推理 = 训练，同一入口\"**——把\"Agent 如何与环境交互并产出 token-level 训练数据\"封装成一套可复用、可扩展、可训练的工程系统。无论你是跑 1000 个并行的 SWE-Bench 推理任务，还是进行 GRPO/GSPO 强化学习训练，底层走的是同一套 `AgentInteraction` 交互循环。这种设计使得从\"模型推理验证\"到\"大规模 RL 训练\"的切换成本趋近于零——只需换一个 YAML 配置，不需要重写任何交互逻辑。", "url": "https://www.cnblogs.com/rossiXYZ/p/20431463", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20431463", "series": "uni-agent", "seriesPart": 2, "space": "deep", "coverImage": "/assets/covers/uni-agent-02-key-tech.jpg", "items": [ { "title": "Uni-Agent 系列", "titleZh": "📚 Uni-Agent 系列", "summary": "共 2 篇 · 站内专栏", "category": "专栏", "seriesKey": "uni-agent" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 修改扩展点", "titleZh": "0x01 修改扩展点", "category": "章节", "url": "" }, { "title": "0x02 Uni-Agent 关键技术", "titleZh": "0x02 Uni-Agent 关键技术", "category": "章节", "url": "" }, { "title": "0x03 Agentic RL 八维度评估", "titleZh": "0x03 Agentic RL 八维度评估", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20431463", "external": true } ], "bodyPath": "/data/bodies/uni-agent-02-key-tech.json" }, { "id": "uni-agent-01-overview", "title": "【Agentic RL / 强化学习框架】Uni-Agent 深度技术分析（1）--- 总体", "category": "paper", "categoryName": "AI 论文", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-06-09", "likes": 0, "featured": true, "summary": "项目的核心主张可以概括为 **\"推理 = 训练，同一入口\"**——把\"Agent 如何与环境交互并产出 token-level 训练数据\"封装成一套可复用、可扩展、可训练的工程系统。无论你是跑 1000 个并行的 SWE-Bench 推理任务，还是进行 GRPO/GSPO 强化学习训练，底层走的是同一套 `AgentInteraction…", "overview": "项目的核心主张可以概括为 **\"推理 = 训练，同一入口\"**——把\"Agent 如何与环境交互并产出 token-level 训练数据\"封装成一套可复用、可扩展、可训练的工程系统。无论你是跑 1000 个并行的 SWE-Bench 推理任务，还是进行 GRPO/GSPO 强化学习训练，底层走的是同一套 `AgentInteraction` 交互循环。这种设计使得从\"模型推理验证\"到\"大规模 RL 训练\"的切换成本趋近于零——只需换一个 YAML 配置，不需要重写任何交互逻辑。", "url": "https://www.cnblogs.com/rossiXYZ/p/20407294", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20407294", "series": "uni-agent", "seriesPart": 1, "space": "deep", "coverImage": "/assets/covers/uni-agent-01-overview.jpg", "items": [ { "title": "Uni-Agent 系列", "titleZh": "📚 Uni-Agent 系列", "summary": "共 2 篇 · 站内专栏", "category": "专栏", "seriesKey": "uni-agent" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 基本功能", "titleZh": "0x01 基本功能", "category": "章节", "url": "" }, { "title": "0x02 Agentic RL 的难点", "titleZh": "0x02 Agentic RL 的难点", "category": "章节", "url": "" }, { "title": "0x03 verl 的能力与系统性不足", "titleZh": "0x03 verl 的能力与系统性不足", "category": "章节", "url": "" }, { "title": "0x04 Uni-Agent 设计理念与核心功能", "titleZh": "0x04 Uni-Agent 设计理念与核心功能", "category": "章节", "url": "" }, { "title": "0x05 Uni-Agent 对 verl 的改造", "titleZh": "0x05 Uni-Agent 对 verl 的改造", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20407294", "external": true } ], "bodyPath": "/data/bodies/uni-agent-01-overview.json" }, { "id": "edge-agent-os-phone", "title": "端侧 Agent OS：手机架构与硬件协同设计", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-06-08", "likes": 0, "featured": false, "summary": "在进入具体硬件和软件设计之前，必须先回答：端侧 Agent OS 到底和云侧有什么 **本质** 差异？这些差异不是程度上的（\"资源少一点\"），而是本质上的——它们会反向塑造整个架构。", "overview": "在进入具体硬件和软件设计之前，必须先回答：端侧 Agent OS 到底和云侧有什么 **本质** 差异？这些差异不是程度上的（\"资源少一点\"），而是本质上的——它们会反向塑造整个架构。", "url": "https://www.cnblogs.com/rossiXYZ/p/20350650", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20350650", "series": null, "seriesPart": 0, "space": "deep", "coverImage": "/assets/covers/edge-agent-os-phone.jpg", "items": [ { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 Part 1：核心论断 — 确定性基线与端侧定位", "titleZh": "0x01 Part 1：核心论断 — 确定性基线与端侧定位", "category": "章节", "url": "" }, { "title": "0x02 Part 2：硬件基础 — SoC 异构与限制", "titleZh": "0x02 Part 2：硬件基础 — SoC 异构与限制", "category": "章节", "url": "" }, { "title": "0x03 Part 3：软件架构 — OS 适配与端侧 Harness", "titleZh": "0x03 Part 3：软件架构 — OS 适配与端侧 Harness", "category": "章节", "url": "" }, { "title": "0x04 Part 4：范式特化 — 五种范式 × 确定性基线", "titleZh": "0x04 Part 4：范式特化 — 五种范式 × 确定性基线", "category": "章节", "url": "" }, { "title": "0x05 Part 5：安全与信任 — TEE + Agent 状态可见性", "titleZh": "0x05 Part 5：安全与信任 — TEE + Agent 状态可见性", "category": "章节", "url": "" }, { "title": "0x06 Part 6：案例", "titleZh": "0x06 Part 6：案例", "category": "章节", "url": "" }, { "title": "0x07 Part 7：战略定位", "titleZh": "0x07 Part 7：战略定位", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20350650", "external": true } ], "bodyPath": "/data/bodies/edge-agent-os-phone.json" }, { "id": "agent-os-five-paradigms", "title": "Agent OS ：五种驯服不确定性的范式", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-06-04", "likes": 0, "featured": false, "summary": "可能是3年前，那时候很多人都在说AIOS，但是到了今天，恐怕Agent OS这个名词更符合目前的趋势。", "overview": "可能是3年前，那时候很多人都在说AIOS，但是到了今天，恐怕Agent OS这个名词更符合目前的趋势。", "url": "https://www.cnblogs.com/rossiXYZ/p/20296163", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20296163", "series": null, "seriesPart": 0, "space": "deep", "coverImage": "/assets/covers/agent-os-five-paradigms.jpg", "items": [ { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 Part 1: 问题空间", "titleZh": "0x01 Part 1: 问题空间", "category": "章节", "url": "" }, { "title": "0x02 Part 2: 理论基础", "titleZh": "0x02 Part 2: 理论基础", "category": "章节", "url": "" }, { "title": "0x03 Part 3: 状态基石", "titleZh": "0x03 Part 3: 状态基石", "category": "章节", "url": "" }, { "title": "0x04 Part 4: 七层架构 — ETCLOVG 详解", "titleZh": "0x04 Part 4: 七层架构 — ETCLOVG 详解", "category": "章节", "url": "" }, { "title": "0x05 Part 5: 工程实践", "titleZh": "0x05 Part 5: 工程实践", "category": "章节", "url": "" }, { "title": "0x06 Part 6：总结与展望", "titleZh": "0x06 Part 6：总结与展望", "category": "章节", "url": "" }, { "title": "0xEE 广告", "titleZh": "0xEE 广告", "category": "章节", "url": "" }, { "title": "0xFF 参考/知识来源", "titleZh": "0xFF 参考/知识来源", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20296163", "external": true } ], "bodyPath": "/data/bodies/agent-os-five-paradigms.json" }, { "id": "agent-17-architecture-patterns", "title": "Agent 17 种架构模式分析 & 思考", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-06-02", "likes": 0, "featured": false, "summary": "**注**：太卷了太卷了。我还没看完 miles，朱大佬就放出了面向 Agent 时代的 slime v0.3.0。我写博客/写PPT的速度都比不上业界开发的速度。", "overview": "**注**：太卷了太卷了。我还没看完 miles，朱大佬就放出了面向 Agent 时代的 slime v0.3.0。我写博客/写PPT的速度都比不上业界开发的速度。", "url": "https://www.cnblogs.com/rossiXYZ/p/20274878", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20274878", "series": null, "seriesPart": 17, "space": "deep", "coverImage": "/assets/covers/agent-17-architecture-patterns.jpg", "items": [ { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 核心概念", "titleZh": "0x01 核心概念", "category": "章节", "url": "" }, { "title": "0x02 评估维度", "titleZh": "0x02 评估维度", "category": "章节", "url": "" }, { "title": "0x03 17 种架构总述", "titleZh": "0x03 17 种架构总述", "category": "章节", "url": "" }, { "title": "0x04 17 种架构逐一分析", "titleZh": "0x04 17 种架构逐一分析", "category": "章节", "url": "" }, { "title": "0x05 演化路径与升级触发条件", "titleZh": "0x05 演化路径与升级触发条件", "category": "章节", "url": "" }, { "title": "0xEE 广告", "titleZh": "0xEE 广告", "category": "章节", "url": "" }, { "title": "0xFF 参考来源", "titleZh": "0xFF 参考来源", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20274878", "external": true } ], "bodyPath": "/data/bodies/agent-17-architecture-patterns.json" }, { "id": "openclaw-rl-05-async", "title": "【Agentic RL / 强化学习 / OPD】OpenClaw-RL 源码阅读笔记 --- (5)--- 异步处理", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-06-01", "likes": 0, "featured": false, "summary": "本系列的目的是：借着对 OpenClaw-RL 源码的学习，来梳理强化学习的一些相关概念和思想。所以，会有一些基础知识、扩展和发散，OpenClaw-RL 只是一个切入点。而且，因为整篇系列是一个整体，所以有些概念的解读/学习会在不同的文章中出现，还请大家谅解。", "overview": "本系列的目的是：借着对 OpenClaw-RL 源码的学习，来梳理强化学习的一些相关概念和思想。所以，会有一些基础知识、扩展和发散，OpenClaw-RL 只是一个切入点。而且，因为整篇系列是一个整体，所以有些概念的解读/学习会在不同的文章中出现，还请大家谅解。", "url": "https://www.cnblogs.com/rossiXYZ/p/20240844", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20240844", "series": "openclaw-rl", "seriesPart": 5, "space": "deep", "coverImage": "/assets/covers/openclaw-rl-05-async.jpg", "items": [ { "title": "OpenClaw-RL 源码系列", "titleZh": "📚 OpenClaw-RL 源码系列", "summary": "共 7 篇 · 站内专栏", "category": "专栏", "seriesKey": "openclaw-rl" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 异步架构", "titleZh": "0x01 异步架构", "category": "章节", "url": "" }, { "title": "0x02 训推分离", "titleZh": "0x02 训推分离", "category": "章节", "url": "" }, { "title": "0x03 调度策略", "titleZh": "0x03 调度策略", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20240844", "external": true } ], "bodyPath": "/data/bodies/openclaw-rl-05-async.json" }, { "id": "openclaw-rl-04-arch", "title": "【Agentic RL / 强化学习 / OPD】OpenClaw-RL 源码阅读笔记 --- (4)--- 架构", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-05-30", "likes": 0, "featured": false, "summary": "本系列的目的是：借着对 OpenClaw-RL 源码的学习，来梳理强化学习的一些相关概念和思想。所以，会有一些基础知识、扩展和发散，OpenClaw-RL 只是一个切入点。而且，因为整篇系列是一个整体，所以有些概念的解读/学习会在不同的文章中出现，还请大家谅解。", "overview": "本系列的目的是：借着对 OpenClaw-RL 源码的学习，来梳理强化学习的一些相关概念和思想。所以，会有一些基础知识、扩展和发散，OpenClaw-RL 只是一个切入点。而且，因为整篇系列是一个整体，所以有些概念的解读/学习会在不同的文章中出现，还请大家谅解。", "url": "https://www.cnblogs.com/rossiXYZ/p/20229957", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20229957", "series": "openclaw-rl", "seriesPart": 4, "space": "deep", "coverImage": "/assets/covers/openclaw-rl-04-arch.jpg", "items": [ { "title": "OpenClaw-RL 源码系列", "titleZh": "📚 OpenClaw-RL 源码系列", "summary": "共 7 篇 · 站内专栏", "category": "专栏", "seriesKey": "openclaw-rl" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 架构", "titleZh": "0x01 架构", "category": "章节", "url": "" }, { "title": "0x02 Slime 的作用", "titleZh": "0x02 Slime 的作用", "category": "章节", "url": "" }, { "title": "0x03 交互", "titleZh": "0x03 交互", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20229957", "external": true } ], "bodyPath": "/data/bodies/openclaw-rl-04-arch.json" }, { "id": "openclaw-rl-03-overview", "title": "【Agentic RL / 强化学习 / OPD】OpenClaw-RL 源码阅读笔记 --- (3)--- 总体思考", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-05-28", "likes": 0, "featured": false, "summary": "本系列的目的是：借着对 OpenClaw-RL 源码的学习，来梳理强化学习的一些相关概念和思想。所以，会有一些基础概念、扩展和发散，OpenClaw-RL 只是一个切入点。而且，因为整篇系列是一个整体，所以有些概念的解读/学习会在不同的文章中出现，还请大家谅解。", "overview": "本系列的目的是：借着对 OpenClaw-RL 源码的学习，来梳理强化学习的一些相关概念和思想。所以，会有一些基础概念、扩展和发散，OpenClaw-RL 只是一个切入点。而且，因为整篇系列是一个整体，所以有些概念的解读/学习会在不同的文章中出现，还请大家谅解。", "url": "https://www.cnblogs.com/rossiXYZ/p/20141000", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20141000", "series": "openclaw-rl", "seriesPart": 3, "space": "deep", "coverImage": "/assets/covers/openclaw-rl-03-overview.jpg", "items": [ { "title": "OpenClaw-RL 源码系列", "titleZh": "📚 OpenClaw-RL 源码系列", "summary": "共 7 篇 · 站内专栏", "category": "专栏", "seriesKey": "openclaw-rl" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 四个要点", "titleZh": "0x01 四个要点", "category": "章节", "url": "" }, { "title": "0x02 第一不变量理解", "titleZh": "0x02 第一不变量理解", "category": "章节", "url": "" }, { "title": "0x03 第二变量理解", "titleZh": "0x03 第二变量理解", "category": "章节", "url": "" }, { "title": "0xEE 广告", "titleZh": "0xEE 广告", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20141000", "external": true } ], "bodyPath": "/data/bodies/openclaw-rl-03-overview.json" }, { "id": "openclaw-rl-02-opd", "title": "OpenClaw-RL (2)— On-Policy Distillation", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-05-27", "likes": 0, "featured": false, "summary": "本系列的目的是：借着对 OpenClaw-RL 源码的学习，来梳理强化学习的一些相关概念和思想。所以，会有一些基础知识、扩展和发散，OpenClaw-RL 只是一个切入点。而且，因为整篇系列是一个整体，所以有些概念的解读/学习会在不同的文章中出现，还请大家谅解。", "overview": "本系列的目的是：借着对 OpenClaw-RL 源码的学习，来梳理强化学习的一些相关概念和思想。所以，会有一些基础知识、扩展和发散，OpenClaw-RL 只是一个切入点。而且，因为整篇系列是一个整体，所以有些概念的解读/学习会在不同的文章中出现，还请大家谅解。", "url": "https://www.cnblogs.com/rossiXYZ/p/20140091", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20140091", "series": "openclaw-rl", "seriesPart": 2, "space": "deep", "coverImage": "/assets/covers/openclaw-rl-02-opd.jpg", "items": [ { "title": "OpenClaw-RL 源码系列", "titleZh": "📚 OpenClaw-RL 源码系列", "summary": "共 7 篇 · 站内专栏", "category": "专栏", "seriesKey": "openclaw-rl" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 On-Policy Distillation", "titleZh": "0x01 On-Policy Distillation", "category": "章节", "url": "" }, { "title": "0x02 Hindsight-Guided On-Policy Distillation", "titleZh": "0x02 Hindsight-Guided On-Policy Distillation", "category": "章节", "url": "" }, { "title": "0x03 对比", "titleZh": "0x03 对比", "category": "章节", "url": "" }, { "title": "0xEE 广告", "titleZh": "0xEE 广告", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20140091", "external": true } ], "bodyPath": "/data/bodies/openclaw-rl-02-opd.json" }, { "id": "openclaw-rl-01-basics", "title": "【Agentic RL / 强化学习 / OPD】OpenClaw-RL 源码阅读笔记 --- (1)---基础", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-05-25", "likes": 0, "featured": true, "summary": "本系列的目的是：借着对 OpenClaw-RL 源码的学习，来梳理强化学习的一些相关概念和思想。所以，会有一些扩展和发散，OpenClaw-RL 只是一个切入点。而且，因为整篇系列是一个整体，所以有些概念的解读/学习会在不同的文章中出现，还请大家谅解。", "overview": "本系列的目的是：借着对 OpenClaw-RL 源码的学习，来梳理强化学习的一些相关概念和思想。所以，会有一些扩展和发散，OpenClaw-RL 只是一个切入点。而且，因为整篇系列是一个整体，所以有些概念的解读/学习会在不同的文章中出现，还请大家谅解。", "url": "https://www.cnblogs.com/rossiXYZ/p/20137685", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20137685", "series": "openclaw-rl", "seriesPart": 1, "space": "deep", "coverImage": "/assets/covers/openclaw-rl-01-basics.jpg", "items": [ { "title": "OpenClaw-RL 源码系列", "titleZh": "📚 OpenClaw-RL 源码系列", "summary": "共 7 篇 · 站内专栏", "category": "专栏", "seriesKey": "openclaw-rl" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 背景知识", "titleZh": "0x01 背景知识", "category": "章节", "url": "" }, { "title": "0x02 论文基础", "titleZh": "0x02 论文基础", "category": "章节", "url": "" }, { "title": "0x03 Roadmap", "titleZh": "0x03 Roadmap", "category": "章节", "url": "" }, { "title": "0xEE 广告", "titleZh": "0xEE 广告", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20137685", "external": true } ], "bodyPath": "/data/bodies/openclaw-rl-01-basics.json" }, { "id": "kernelagent-06-composer", "title": "PyTorch KernelAgent 源码解读 ---（6）--- Composer", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-05-22", "likes": 0, "featured": false, "summary": "如果 LLM 生成的第一个组合内核无法通过验证（运行 / 编译失败），该脚本会捕获错误信息（`stderr`，`stdout`）。它会构建一个新的提示（`_build_refinement_prompt`），将错误信息作为上下文提供给 LLM，要求其修正代码。这个过程可以重复多次（`max_iters` 参数控制最大迭代次数），直到生成的…", "overview": "· **子图内核整合**：将 Fuser 流程中拆分的子图及其对应的、已验证的 Triton 内核，重新组合为一个端到端的 Triton 实现，替代原始 PyTorch 代码的前向传播逻辑。\n· **LLM 驱动的代码生成**：以原始问题代码、子图信息、子图 Triton 内核为输入，通过定制化 Prompt 调用 LLM 生成完整的 Triton 内核代码。\n· **功能验证与迭代优化**：支持自动验证生成的内核（对比 PyTorch 参考结果），若验证失败则基于错误信息迭代调用 LLM 修正代码，直到通过验证或达到最大迭代次数。\n· **约束保障**：通过严格的代码规范（如必须包含 `kernel_function`、禁止 PyTorch 计算逻辑）和数值校验，确保生成的 Triton 内核可用、正确。", "url": "https://www.cnblogs.com/rossiXYZ/p/20124922", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20124922", "series": "kernelagent", "seriesPart": 6, "space": "deep", "coverImage": "/assets/covers/kernelagent-06-composer.jpg", "items": [ { "title": "KernelAgent 源码系列", "titleZh": "📚 KernelAgent 源码系列", "summary": "共 6 篇 · 站内专栏", "category": "专栏", "seriesKey": "kernelagent" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 核心功能", "titleZh": "0x01 核心功能", "category": "章节", "url": "" }, { "title": "0x02 详细功能", "titleZh": "0x02 详细功能", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20124922", "external": true } ], "bodyPath": "/data/bodies/kernelagent-06-composer.json" }, { "id": "kernelagent-05-dispatcher", "title": "PyTorch KernelAgent 源码解读 ---（5）--- Dispatcher", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-05-20", "likes": 0, "featured": false, "summary": "dispatch\\_kernel\\_agent.py 是 KernelAgent 系统中的调度组件，负责将 subgraph\\_extractor.py 生成的子图（JSON 格式）转换为具体的 Triton 内核生成任务，并调度 TritonKernelAgent 来生成和验证这些内核。", "overview": "dispatch\\_kernel\\_agent.py 是 KernelAgent 系统中的调度组件，负责将 subgraph\\_extractor.py 生成的子图（JSON 格式）转换为具体的 Triton 内核生成任务，并调度 TritonKernelAgent 来生成和验证这些内核。", "url": "https://www.cnblogs.com/rossiXYZ/p/20099042", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20099042", "series": "kernelagent", "seriesPart": 5, "space": "deep", "coverImage": "/assets/covers/kernelagent-05-dispatcher.jpg", "items": [ { "title": "KernelAgent 源码系列", "titleZh": "📚 KernelAgent 源码系列", "summary": "共 6 篇 · 站内专栏", "category": "专栏", "seriesKey": "kernelagent" }, { "title": "0x00 概述", "titleZh": "0x00 概述", "category": "章节", "url": "" }, { "title": "0x01 Dispatch Kernel Agent 的作用", "titleZh": "0x01 Dispatch Kernel Agent 的作用", "category": "章节", "url": "" }, { "title": "0x02 TritonKernelAgent", "titleZh": "0x02 TritonKernelAgent", "category": "章节", "url": "" }, { "title": "0x03 prompt", "titleZh": "0x03 prompt", "category": "章节", "url": "" }, { "title": "0x04 Worker Manager 的作用", "titleZh": "0x04 Worker Manager 的作用", "category": "章节", "url": "" }, { "title": "0x05 VerificationWorker", "titleZh": "0x05 VerificationWorker", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20099042", "external": true } ], "bodyPath": "/data/bodies/kernelagent-05-dispatcher.json" }, { "id": "kernelagent-04-extractor", "title": "PyTorch KernelAgent 源码解读 ---（4）--- ExtractorAgent", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-05-18", "likes": 0, "featured": false, "summary": "subgraph\\_extractor.py 是子图提取的关键组件，用于从融合的 PyTorch 代码中识别和提取可融合的子图及其形状信息。这个组件使得复杂的融合操作可以分解为更小、更易管理的子图，每个子图都有精确的形状和语义信息，为后续的 Triton 内核生成奠定了坚实基础。", "overview": "subgraph\\_extractor.py 是子图提取的关键组件，用于从融合的 PyTorch 代码中识别和提取可融合的子图及其形状信息。这个组件使得复杂的融合操作可以分解为更小、更易管理的子图，每个子图都有精确的形状和语义信息，为后续的 Triton 内核生成奠定了坚实基础。", "url": "https://www.cnblogs.com/rossiXYZ/p/20065909", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20065909", "series": "kernelagent", "seriesPart": 4, "space": "deep", "coverImage": "/assets/covers/kernelagent-04-extractor.jpg", "items": [ { "title": "KernelAgent 源码系列", "titleZh": "📚 KernelAgent 源码系列", "summary": "共 6 篇 · 站内专栏", "category": "专栏", "seriesKey": "kernelagent" }, { "title": "0x00 摘要", "titleZh": "0x00 摘要", "category": "章节", "url": "" }, { "title": "0x01 功能详解", "titleZh": "0x01 功能详解", "category": "章节", "url": "" }, { "title": "0x02 Prompt", "titleZh": "0x02 Prompt", "category": "章节", "url": "" }, { "title": "0x03 实现", "titleZh": "0x03 实现", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20065909", "external": true } ], "bodyPath": "/data/bodies/kernelagent-04-extractor.json" }, { "id": "kernelagent-03-orchestrator", "title": "PyTorch KernelAgent 源码解读 ---（3）--- orchestrator", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-05-16", "likes": 0, "featured": false, "summary": "orchestrator 是 KernelAgent 系统中的一个核心组件，负责协调和管理多个工作进程(worker)，实现并行执行任务并从中选择最优结果。Fuser/orchestrator.py 文件实现了 Orchestrator 类，用于多进程协调任务执行。其功能用一句话概括：fork N 个Worker竞赛，首个 PASS胜出，…", "overview": "orchestrator 是 KernelAgent 系统中的一个核心组件，负责协调和管理多个工作进程(worker)，实现并行执行任务并从中选择最优结果。Fuser/orchestrator.py 文件实现了 Orchestrator 类，用于多进程协调任务执行。其功能用一句话概括：fork N 个Worker竞赛，首个 PASS胜出，其余终止，产物打包返回。", "url": "https://www.cnblogs.com/rossiXYZ/p/20061518", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20061518", "series": "kernelagent", "seriesPart": 3, "space": "deep", "coverImage": "/assets/covers/kernelagent-03-orchestrator.jpg", "items": [ { "title": "KernelAgent 源码系列", "titleZh": "📚 KernelAgent 源码系列", "summary": "共 6 篇 · 站内专栏", "category": "专栏", "seriesKey": "kernelagent" }, { "title": "0x00 摘要", "titleZh": "0x00 摘要", "category": "章节", "url": "" }, { "title": "0x01 Fuse", "titleZh": "0x01 Fuse", "category": "章节", "url": "" }, { "title": "0x02 多 Worker 并行机制详解", "titleZh": "0x02 多 Worker 并行机制详解", "category": "章节", "url": "" }, { "title": "0x03 KernelAgent 中的 Prompt 特点分析", "titleZh": "0x03 KernelAgent 中的 Prompt 特点分析", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20061518", "external": true } ], "bodyPath": "/data/bodies/kernelagent-03-orchestrator.json" }, { "id": "kernelagent-02-pipeline", "title": "PyTorch KernelAgent 源码解读 ---（2）--- 总体流程", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-05-13", "likes": 0, "featured": false, "summary": "KernelFalcon 是PyTorch 提出的一个Deep Agents架构系统，该框架主要尝试利用Agent端到端实现torch模型优化及 Triton算子自动生成，是首个在全部 250 个 L1/L2/L3 KernelBench 任务上达到 100% 正确率的开源智能体系统。", "overview": "KernelFalcon 是PyTorch 提出的一个Deep Agents架构系统，该框架主要尝试利用Agent端到端实现torch模型优化及 Triton算子自动生成，是首个在全部 250 个 L1/L2/L3 KernelBench 任务上达到 100% 正确率的开源智能体系统。", "url": "https://www.cnblogs.com/rossiXYZ/p/20009243", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "20009243", "series": "kernelagent", "seriesPart": 2, "space": "deep", "coverImage": "/assets/covers/kernelagent-02-pipeline.jpg", "items": [ { "title": "KernelAgent 源码系列", "titleZh": "📚 KernelAgent 源码系列", "summary": "共 6 篇 · 站内专栏", "category": "专栏", "seriesKey": "kernelagent" }, { "title": "0x00 摘要", "titleZh": "0x00 摘要", "category": "章节", "url": "" }, { "title": "0x01 引言", "titleZh": "0x01 引言", "category": "章节", "url": "" }, { "title": "0x02 架构", "titleZh": "0x02 架构", "category": "章节", "url": "" }, { "title": "0x03 架构：分阶段详解", "titleZh": "0x03 架构：分阶段详解", "category": "章节", "url": "" }, { "title": "0x04 KernelAgent 项目入口点", "titleZh": "0x04 KernelAgent 项目入口点", "category": "章节", "url": "" }, { "title": "0x05 pipeline.py 的作用分析", "titleZh": "0x05 pipeline.py 的作用分析", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/20009243", "external": true } ], "bodyPath": "/data/bodies/kernelagent-02-pipeline.json" }, { "id": "kernelagent-01-principles", "title": "PyTorch KernelAgent 源码解读 ---（1）--- 原理", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-05-11", "likes": 0, "featured": true, "summary": "在深度学习模型的推理与训练过程中，绝大部分计算都依赖于底层计算内核（Kernel）来执行。计算内核是运行在硬件加速器（如 GPU、NPU、TPU）上的 “小型高性能程序”，它负责完成矩阵乘法、卷积、归一化等深度学习的核心算子运算。", "overview": "· 模型侧：大模型（LLM）领域，稠密模型、MOE、MLA、多模态场景众多；稀疏、量化、KVCache压缩等技术又为算子生成带来更多复杂性与多样性。推荐类、 CV类模型同样在持续演进，与LLM截然不同的算子场景同样存在强烈的优化落地需求；\n· 硬件侧：Nvidia GPU的不断升级，国产AI芯片的持续不断演进，硬件架构、特性、参数各有不同；为提高AI芯片可用性，构建性能优势，定制化算子优化方案需求强烈；", "url": "https://www.cnblogs.com/rossiXYZ/p/19963949", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19963949", "series": "kernelagent", "seriesPart": 1, "space": "deep", "coverImage": "/assets/covers/kernelagent-01-principles.jpg", "items": [ { "title": "KernelAgent 源码系列", "titleZh": "📚 KernelAgent 源码系列", "summary": "共 6 篇 · 站内专栏", "category": "专栏", "seriesKey": "kernelagent" }, { "title": "0x00 摘要", "titleZh": "0x00 摘要", "category": "章节", "url": "" }, { "title": "0x01 背景知识", "titleZh": "0x01 背景知识", "category": "章节", "url": "" }, { "title": "0x02 微调范式", "titleZh": "0x02 微调范式", "category": "章节", "url": "" }, { "title": "0x03 Agent Pipeline 范式", "titleZh": "0x03 Agent Pipeline 范式", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19963949", "external": true } ], "bodyPath": "/data/bodies/kernelagent-01-principles.json" }, { "id": "mai-ui-02-impl", "title": "【GUI-Agent】阿里通义MAI-UI 代码阅读（2）--- 实现", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-05-08", "likes": 0, "featured": false, "summary": "MAI-UI 是阿里通义实验室发布的一项重磅研究成果：是一个旨在 **重塑人机交互方式** 的“ **基础图形用户界面（GUI）智能体**”，和阶跃星辰的思路非常类似，因此我们可以互相印证。", "overview": "MAI-UI 是阿里通义实验室发布的一项重磅研究成果：是一个旨在 **重塑人机交互方式** 的“ **基础图形用户界面（GUI）智能体**”，和阶跃星辰的思路非常类似，因此我们可以互相印证。", "url": "https://www.cnblogs.com/rossiXYZ/p/19963833", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19963833", "series": "mai-ui", "seriesPart": 2, "space": "deep", "coverImage": "/assets/covers/mai-ui-02-impl.jpg", "items": [ { "title": "MAI-UI 系列", "titleZh": "📚 MAI-UI 系列", "summary": "共 2 篇 · 站内专栏", "category": "专栏", "seriesKey": "mai-ui" }, { "title": "0x00 摘要", "titleZh": "0x00 摘要", "category": "章节", "url": "" }, { "title": "0x01 工程实现特色", "titleZh": "0x01 工程实现特色", "category": "章节", "url": "" }, { "title": "0x02 提示词", "titleZh": "0x02 提示词", "category": "章节", "url": "" }, { "title": "0x03 输出", "titleZh": "0x03 输出", "category": "章节", "url": "" }, { "title": "0x04 MAIUINaivigationAgent", "titleZh": "0x04 MAIUINaivigationAgent", "category": "章节", "url": "" }, { "title": "0x05 MAIGroundingAgent", "titleZh": "0x05 MAIGroundingAgent", "category": "章节", "url": "" }, { "title": "0xEE 广告", "titleZh": "0xEE 广告", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19963833", "external": true } ], "bodyPath": "/data/bodies/mai-ui-02-impl.json" }, { "id": "mai-ui-01-overview", "title": "【GUI-Agent】阿里通义MAI-UI 代码阅读（1）--- 总体", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-05-06", "likes": 0, "featured": true, "summary": "MAI-UI是阿里通义实验室发布的一项重磅研究成果：是一个旨在 **重塑人机交互方式** 的“ **基础图形用户界面（GUI）智能体**”，和阶跃星辰的思路非常类似，因此我们可以互相印证。", "overview": "MAI-UI是阿里通义实验室发布的一项重磅研究成果：是一个旨在 **重塑人机交互方式** 的“ **基础图形用户界面（GUI）智能体**”，和阶跃星辰的思路非常类似，因此我们可以互相印证。", "url": "https://www.cnblogs.com/rossiXYZ/p/19963830", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19963830", "series": "mai-ui", "seriesPart": 1, "space": "deep", "coverImage": "/assets/covers/mai-ui-01-overview.jpg", "items": [ { "title": "MAI-UI 系列", "titleZh": "📚 MAI-UI 系列", "summary": "共 2 篇 · 站内专栏", "category": "专栏", "seriesKey": "mai-ui" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19963830", "external": true } ], "bodyPath": "/data/bodies/mai-ui-01-overview.json" }, { "id": "nanobot-10-heartbeat", "title": "【OpenClaw】通过 Nanobot 源码学习架构---（10）Heartbeat", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-04-27", "likes": 0, "featured": false, "summary": "HEARTBEAT.md是一个标记文件，包含需要定期检查的任务列表，文件位于工作空间根目录，可由agent自主更新。agent 可以使用文件工具（如 edit\\_file、write\\_file）更新 HEARTBEAT.md，支持添加、移除或重写周期性任务。", "overview": "· 按配置的时间间隔（默认 30 分钟）自动唤醒，读取工作目录下的`HEARTBEAT.md`文件，并执行 HEARTBEAT.md 中的周期性任务；\n· 两阶段执行模式：`HeartbeatService`采用 “两阶段执行” 架构，将 “决策” 和 “执行” 解耦，既保证决策的智能化，又实现执行逻辑的解耦：\n· LLM 驱动的智能决策：放弃传统的 “关键字匹配 / 正则解析” 方式，通过 LLM + 虚拟工具调用的方式分析`HEARTBEAT.md`内容，判断是否有任务需要执行，避免了 “HEARTBEAT\\_OK” 这类硬编码令牌的不稳定性，适配自然语言描述的任务场景。\n· 灵活的回调执行扩展：通过`on_execute`和`on_notify`回调函数解耦 “任务执行” 和 “结果推送” 逻辑，无需修改心跳服务核心代码即可适配不同的执行 / 推送策略。", "url": "https://www.cnblogs.com/rossiXYZ/p/19852143", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19852143", "series": "nanobot", "seriesPart": 10, "space": "deep", "coverImage": "/assets/covers/nanobot-10-heartbeat.jpg", "items": [ { "title": "Nanobot 架构系列", "titleZh": "📚 Nanobot 架构系列", "summary": "共 10 篇 · 站内专栏", "category": "专栏", "seriesKey": "nanobot" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 基本功能", "titleZh": "0x01 基本功能", "category": "章节", "url": "" }, { "title": "0x02 详细分析", "titleZh": "0x02 详细分析", "category": "章节", "url": "" }, { "title": "0x03 对比", "titleZh": "0x03 对比", "category": "章节", "url": "" }, { "title": "0x04 HeartbeatService实现", "titleZh": "0x04 HeartbeatService实现", "category": "章节", "url": "" }, { "title": "0xEE 广告", "titleZh": "0xEE 广告", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19852143", "external": true } ], "bodyPath": "/data/bodies/nanobot-10-heartbeat.json" }, { "id": "nanobot-09-periodic", "title": "【OpenClaw】通过 Nanobot 源码学习架构---（9）周期性执行", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-04-24", "likes": 0, "featured": false, "summary": "Nanobot项目通过 CronService 和 CronTool 实现周期性任务：CronTool 作为面向 LLM 的接口，CronService 作为实际的业务逻辑执行者，两者通过依赖注入的方式协作，形成了一个完整的定时任务管理解决方案。", "overview": "Nanobot项目通过 CronService 和 CronTool 实现周期性任务：CronTool 作为面向 LLM 的接口，CronService 作为实际的业务逻辑执行者，两者通过依赖注入的方式协作，形成了一个完整的定时任务管理解决方案。", "url": "https://www.cnblogs.com/rossiXYZ/p/19852068", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19852068", "series": "nanobot", "seriesPart": 9, "space": "deep", "coverImage": "/assets/covers/nanobot-09-periodic.jpg", "items": [ { "title": "Nanobot 架构系列", "titleZh": "📚 Nanobot 架构系列", "summary": "共 10 篇 · 站内专栏", "category": "专栏", "seriesKey": "nanobot" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 基本知识", "titleZh": "0x01 基本知识", "category": "章节", "url": "" }, { "title": "0x02 SKILL.md", "titleZh": "0x02 SKILL.md", "category": "章节", "url": "" }, { "title": "0x03 核心组件", "titleZh": "0x03 核心组件", "category": "章节", "url": "" }, { "title": "0x04 CronTool类", "titleZh": "0x04 CronTool类", "category": "章节", "url": "" }, { "title": "0x05 CronService", "titleZh": "0x05 CronService", "category": "章节", "url": "" }, { "title": "0x06 Dream机制", "titleZh": "0x06 Dream机制", "category": "章节", "url": "" }, { "title": "0xEE 广告", "titleZh": "0xEE 广告", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19852068", "external": true } ], "bodyPath": "/data/bodies/nanobot-09-periodic.json" }, { "id": "nanobot-08-tools", "title": "【OpenClaw】通过 Nanobot 源码学习架构---（8）Tools", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-04-22", "likes": 0, "featured": false, "summary": "Tools（函数调用）机制的作用，是由应用侧提供一组可调用的函数，在模型推理过程中，由模型决定是否需要调用这些函数、以及调用哪一个、使用什么参数。模型只负责决策，应用程序负责真正执行函数并返回结果。", "overview": "Tools（函数调用）机制的作用，是由应用侧提供一组可调用的函数，在模型推理过程中，由模型决定是否需要调用这些函数、以及调用哪一个、使用什么参数。模型只负责决策，应用程序负责真正执行函数并返回结果。", "url": "https://www.cnblogs.com/rossiXYZ/p/19843427", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19843427", "series": "nanobot", "seriesPart": 8, "space": "deep", "coverImage": "/assets/covers/nanobot-08-tools.jpg", "items": [ { "title": "Nanobot 架构系列", "titleZh": "📚 Nanobot 架构系列", "summary": "共 10 篇 · 站内专栏", "category": "专栏", "seriesKey": "nanobot" }, { "title": "0x00 摘要", "titleZh": "0x00 摘要", "category": "章节", "url": "" }, { "title": "0x01 原理", "titleZh": "0x01 原理", "category": "章节", "url": "" }, { "title": "0x02 核心代码作用与特色总结", "titleZh": "0x02 核心代码作用与特色总结", "category": "章节", "url": "" }, { "title": "0x03 实现", "titleZh": "0x03 实现", "category": "章节", "url": "" }, { "title": "0xEE 广告", "titleZh": "0xEE 广告", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19843427", "external": true } ], "bodyPath": "/data/bodies/nanobot-08-tools.json" }, { "id": "nanobot-07-memory", "title": "【OpenClaw】通过 Nanobot 源码学习架构---（7）Memory", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-04-13", "likes": 0, "featured": true, "summary": "Memory 是构建复杂 Agent 的关键基础。如果仅依赖 LLM 自带的有限上下文窗口作为短期记忆，Agent 将在跨会话中处于 stateless 的状态，难以维持连贯性。同时，缺乏长期记忆机制也限制了 Agent 自主管理和执行持续性任务的能力。引入额外的 Memory 可以让 Agent 做到存储、检索，并利用历史交互与经验，从…", "overview": "Memory 是构建复杂 Agent 的关键基础。如果仅依赖 LLM 自带的有限上下文窗口作为短期记忆，Agent 将在跨会话中处于 stateless 的状态，难以维持连贯性。同时，缺乏长期记忆机制也限制了 Agent 自主管理和执行持续性任务的能力。引入额外的 Memory 可以让 Agent 做到存储、检索，并利用历史交互与经验，从而实现更智能和持续的行为。", "url": "https://www.cnblogs.com/rossiXYZ/p/19843410", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19843410", "series": "nanobot", "seriesPart": 7, "space": "deep", "coverImage": "/assets/covers/nanobot-07-memory.jpg", "items": [ { "title": "Nanobot 架构系列", "titleZh": "📚 Nanobot 架构系列", "summary": "共 10 篇 · 站内专栏", "category": "专栏", "seriesKey": "nanobot" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 OpenClaw", "titleZh": "0x01 OpenClaw", "category": "章节", "url": "" }, { "title": "0x02 会话消息", "titleZh": "0x02 会话消息", "category": "章节", "url": "" }, { "title": "0x03 MemoryStore", "titleZh": "0x03 MemoryStore", "category": "章节", "url": "" }, { "title": "0x04 SKILL.md", "titleZh": "0x04 SKILL.md", "category": "章节", "url": "" }, { "title": "0x05 相关工具", "titleZh": "0x05 相关工具", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19843410", "external": true } ], "bodyPath": "/data/bodies/nanobot-07-memory.json" }, { "id": "nanobot-06-skills", "title": "【OpenClaw】通过 Nanobot 源码学习架构---（6）Skills", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-04-09", "likes": 0, "featured": false, "summary": "Skill（技能）是将特定领域的专业知识、工作流程和最佳实践打包成可复用的指令模块，嵌入到 AI 的上下文中，使 AI 在处理相关任务时能够自动激活并运用这些专业能力。", "overview": "Skill（技能）是将特定领域的专业知识、工作流程和最佳实践打包成可复用的指令模块，嵌入到 AI 的上下文中，使 AI 在处理相关任务时能够自动激活并运用这些专业能力。", "url": "https://www.cnblogs.com/rossiXYZ/p/19824437", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19824437", "series": "nanobot", "seriesPart": 6, "space": "deep", "coverImage": "/assets/covers/nanobot-06-skills.jpg", "items": [ { "title": "Nanobot 架构系列", "titleZh": "📚 Nanobot 架构系列", "summary": "共 10 篇 · 站内专栏", "category": "专栏", "seriesKey": "nanobot" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 原理", "titleZh": "0x01 原理", "category": "章节", "url": "" }, { "title": "0x02 SkillsLoader", "titleZh": "0x02 SkillsLoader", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19824437", "external": true } ], "bodyPath": "/data/bodies/nanobot-06-skills.json" }, { "id": "nanobot-05-context", "title": "【OpenClaw】通过 Nanobot 源码学习架构---（5）Context", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-04-08", "likes": 0, "featured": false, "summary": "丰富的上下文信息是 Agent 有效规划和行动的基础。一个 Agent 在工作时需要访问的”上下文”如下：", "overview": "丰富的上下文信息是 Agent 有效规划和行动的基础。一个 Agent 在工作时需要访问的”上下文”如下：", "url": "https://www.cnblogs.com/rossiXYZ/p/19824337", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19824337", "series": "nanobot", "seriesPart": 5, "space": "deep", "coverImage": "/assets/covers/nanobot-05-context.jpg", "items": [ { "title": "Nanobot 架构系列", "titleZh": "📚 Nanobot 架构系列", "summary": "共 10 篇 · 站内专栏", "category": "专栏", "seriesKey": "nanobot" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 提示词系统", "titleZh": "0x01 提示词系统", "category": "章节", "url": "" }, { "title": "0x02 ContextBuilder 基本功能", "titleZh": "0x02 ContextBuilder 基本功能", "category": "章节", "url": "" }, { "title": "0x03 图例", "titleZh": "0x03 图例", "category": "章节", "url": "" }, { "title": "0x03 重点函数", "titleZh": "0x03 重点函数", "category": "章节", "url": "" }, { "title": "0x04 代码", "titleZh": "0x04 代码", "category": "章节", "url": "" }, { "title": "0x00 参考", "titleZh": "0x00 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19824337", "external": true } ], "bodyPath": "/data/bodies/nanobot-05-context.json" }, { "id": "nanobot-04-subagent", "title": "【OpenClaw】通过 Nanobot 源码学习架构 ---（4）SubAgent", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-04-06", "likes": 0, "featured": false, "summary": "nanobot 的 Subagent 实现是一个简洁但强大的后台任务执行机制。通过复用主 Agent 的 LLM provider 但限制工具集和迭代次数，实现了任务隔离和资源控制。消息总线机制确保子代理结果能够顺利通知主 Agent，最终传达给用户。这种设计使得主 Agent 可以保持专注的对话交互，同时将复杂任务委派给后台子代理执行。", "overview": "nanobot 的 Subagent 实现是一个简洁但强大的后台任务执行机制。通过复用主 Agent 的 LLM provider 但限制工具集和迭代次数，实现了任务隔离和资源控制。消息总线机制确保子代理结果能够顺利通知主 Agent，最终传达给用户。这种设计使得主 Agent 可以保持专注的对话交互，同时将复杂任务委派给后台子代理执行。", "url": "https://www.cnblogs.com/rossiXYZ/p/19824284", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19824284", "series": "nanobot", "seriesPart": 4, "space": "deep", "coverImage": "/assets/covers/nanobot-04-subagent.jpg", "items": [ { "title": "Nanobot 架构系列", "titleZh": "📚 Nanobot 架构系列", "summary": "共 10 篇 · 站内专栏", "category": "专栏", "seriesKey": "nanobot" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 基础背景", "titleZh": "0x01 基础背景", "category": "章节", "url": "" }, { "title": "0x02 Nanobot SubAgent 功能", "titleZh": "0x02 Nanobot SubAgent 功能", "category": "章节", "url": "" }, { "title": "0x03 Subagent 与 Main Agent 的关系分析", "titleZh": "0x03 Subagent 与 Main Agent 的关系分析", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19824284", "external": true } ], "bodyPath": "/data/bodies/nanobot-04-subagent.json" }, { "id": "nanobot-03-agentloop", "title": "【OpenClaw】通过 Nanobot 源码学习架构---（3）AgentLoop", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-04-03", "likes": 0, "featured": false, "summary": "Agent 是“业务执行者”，解决“消息怎样变成模型调用、工具执行和最终回复”。它们具备独立的上下文（与主对话隔离）和可使用的特定工具，并且具备定义明确的角色和方法论。每次 Agent 收到消息时运行的核心推理周期如下：", "overview": "Agent 是“业务执行者”，解决“消息怎样变成模型调用、工具执行和最终回复”。它们具备独立的上下文（与主对话隔离）和可使用的特定工具，并且具备定义明确的角色和方法论。每次 Agent 收到消息时运行的核心推理周期如下：", "url": "https://www.cnblogs.com/rossiXYZ/p/19818826", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19818826", "series": "nanobot", "seriesPart": 3, "space": "deep", "coverImage": "/assets/covers/nanobot-03-agentloop.jpg", "items": [ { "title": "Nanobot 架构系列", "titleZh": "📚 Nanobot 架构系列", "summary": "共 10 篇 · 站内专栏", "category": "专栏", "seriesKey": "nanobot" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 原理", "titleZh": "0x01 原理", "category": "章节", "url": "" }, { "title": "0x02 AgentLoop", "titleZh": "0x02 AgentLoop", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19818826", "external": true } ], "bodyPath": "/data/bodies/nanobot-03-agentloop.json" }, { "id": "nanobot-02-outer-control", "title": "【OpenClaw】通过 Nanobot 源码学习架构---（2）外层控制逻辑", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-04-01", "likes": 0, "featured": false, "summary": "agent()是CLI的核心实现，是nanobotagent命令的具体函数实现，是nanobot系统的交互式对话入口点。", "overview": "· 任务简单，且不需要隔离上下文。\n· 用户希望对多步骤工作流保有手动控制权。\n· 这个操作会产生某些后果，而用户希望在这些后果发生之前，先由用户自己确认。\n· 用户会在特定时机反复执行的工作流程，而用户希望在自己认为合适的那一刻手动启动它。", "url": "https://www.cnblogs.com/rossiXYZ/p/19808709", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19808709", "series": "nanobot", "seriesPart": 2, "space": "deep", "coverImage": "/assets/covers/nanobot-02-outer-control.jpg", "items": [ { "title": "Nanobot 架构系列", "titleZh": "📚 Nanobot 架构系列", "summary": "共 10 篇 · 站内专栏", "category": "专栏", "seriesKey": "nanobot" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 Commands", "titleZh": "0x01 Commands", "category": "章节", "url": "" }, { "title": "0x02 Gateway", "titleZh": "0x02 Gateway", "category": "章节", "url": "" }, { "title": "0x03 Channel", "titleZh": "0x03 Channel", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19808709", "external": true } ], "bodyPath": "/data/bodies/nanobot-02-outer-control.json" }, { "id": "nanobot-01-overview", "title": "【OpenClaw】通过 Nanobot 源码学习架构---（1）总体", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-03-30", "likes": 0, "featured": true, "summary": "**OpenClaw 是 Harness，是面向个人与本地场景的、开箱即用的 Agent Harness 框架**。它不生产模型，而是把模型 “套上马具”，让模型能稳定、安全、自主地在本地执行真实任务。或者说，OpenClaw 是 Agent 中\"不是 AI 的部分\"，而 Agent 的实际\"聪明程度\"完全取决于背后接入的语言模型。", "overview": "**OpenClaw 是 Harness，是面向个人与本地场景的、开箱即用的 Agent Harness 框架**。它不生产模型，而是把模型 “套上马具”，让模型能稳定、安全、自主地在本地执行真实任务。或者说，OpenClaw 是 Agent 中\"不是 AI 的部分\"，而 Agent 的实际\"聪明程度\"完全取决于背后接入的语言模型。", "url": "https://www.cnblogs.com/rossiXYZ/p/19792125", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19792125", "series": "nanobot", "seriesPart": 1, "space": "deep", "coverImage": "/assets/covers/nanobot-01-overview.jpg", "items": [ { "title": "Nanobot 架构系列", "titleZh": "📚 Nanobot 架构系列", "summary": "共 10 篇 · 站内专栏", "category": "专栏", "seriesKey": "nanobot" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 OpenClaw 基础", "titleZh": "0x01 OpenClaw 基础", "category": "章节", "url": "" }, { "title": "0x02 Nanobot 基础", "titleZh": "0x02 Nanobot 基础", "category": "章节", "url": "" }, { "title": "0x03 Nanobot 总体架构", "titleZh": "0x03 Nanobot 总体架构", "category": "章节", "url": "" }, { "title": "0x04 Nanobot 消息分发机制详解", "titleZh": "0x04 Nanobot 消息分发机制详解", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19792125", "external": true } ], "bodyPath": "/data/bodies/nanobot-01-overview.json" }, { "id": "gui-mcp-06-hitl", "title": "【GUI-Agent】阶跃星辰 GUI-MCP 解读---(6)---HITL(Human In The Loop)", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-03-28", "likes": 0, "featured": false, "summary": "25年底，阶跃星辰升级发布了全新的AI Agent系列模型Step-GUI，包括云端模型Step-GUI、首个面向GUI Agent的MCP协议：GUI-MCP（Graphical User Interface - Model Context Protocol），这是首个专为图形用户界面自动化而设计的 MCP 实现，兼顾标准化与隐私保护。", "overview": "25年底，阶跃星辰升级发布了全新的AI Agent系列模型Step-GUI，包括云端模型Step-GUI、首个面向GUI Agent的MCP协议：GUI-MCP（Graphical User Interface - Model Context Protocol），这是首个专为图形用户界面自动化而设计的 MCP 实现，兼顾标准化与隐私保护。", "url": "https://www.cnblogs.com/rossiXYZ/p/19716573", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19716573", "series": "gui-mcp", "seriesPart": 6, "space": "deep", "coverImage": "/assets/covers/gui-mcp-06-hitl.jpg", "items": [ { "title": "GUI-MCP 解读系列", "titleZh": "📚 GUI-MCP 解读系列", "summary": "共 6 篇 · 站内专栏", "category": "专栏", "seriesKey": "gui-mcp" }, { "title": "0x00 摘要", "titleZh": "0x00 摘要", "category": "章节", "url": "" }, { "title": "0x01 HITL", "titleZh": "0x01 HITL", "category": "章节", "url": "" }, { "title": "0x02 MCP流程", "titleZh": "0x02 MCP流程", "category": "章节", "url": "" }, { "title": "0x03 INFO 操作", "titleZh": "0x03 INFO 操作", "category": "章节", "url": "" }, { "title": "0x04 auto", "titleZh": "0x04 auto", "category": "章节", "url": "" }, { "title": "0x05 特殊分析", "titleZh": "0x05 特殊分析", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19716573", "external": true } ], "bodyPath": "/data/bodies/gui-mcp-06-hitl.json" }, { "id": "gui-mcp-05-cmd-map", "title": "【GUI-Agent】阶跃星辰 GUI-MCP 解读---(5)---命令解析和工具映射", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-03-26", "likes": 0, "featured": false, "summary": "25年底，阶跃星辰升级发布了全新的AI Agent系列模型Step-GUI，包括云端模型Step-GUI、首个面向GUI Agent的MCP协议：GUI-MCP（Graphical User Interface - Model Context Protocol），这是首个专为图形用户界面自动化而设计的 MCP 实现，兼顾标准化与隐私保护。", "overview": "25年底，阶跃星辰升级发布了全新的AI Agent系列模型Step-GUI，包括云端模型Step-GUI、首个面向GUI Agent的MCP协议：GUI-MCP（Graphical User Interface - Model Context Protocol），这是首个专为图形用户界面自动化而设计的 MCP 实现，兼顾标准化与隐私保护。", "url": "https://www.cnblogs.com/rossiXYZ/p/19716388", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19716388", "series": "gui-mcp", "seriesPart": 5, "space": "deep", "coverImage": "/assets/covers/gui-mcp-05-cmd-map.jpg", "items": [ { "title": "GUI-MCP 解读系列", "titleZh": "📚 GUI-MCP 解读系列", "summary": "共 6 篇 · 站内专栏", "category": "专栏", "seriesKey": "gui-mcp" }, { "title": "0x00 摘要", "titleZh": "0x00 摘要", "category": "章节", "url": "" }, { "title": "0x01 MCP 命令解析流程", "titleZh": "0x01 MCP 命令解析流程", "category": "章节", "url": "" }, { "title": "0x02 工具注册机制", "titleZh": "0x02 工具注册机制", "category": "章节", "url": "" }, { "title": "0x03 业务逻辑层", "titleZh": "0x03 业务逻辑层", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19716388", "external": true } ], "bodyPath": "/data/bodies/gui-mcp-05-cmd-map.json" }, { "id": "gui-mcp-04-arch", "title": "【GUI-Agent】阶跃星辰 GUI-MCP 解读---(4)---GUI-MCP 整体架构", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-03-24", "likes": 0, "featured": false, "summary": "25年底，阶跃星辰升级发布了全新的AI Agent系列模型Step-GUI，包括云端模型Step-GUI、首个面向GUI Agent的MCP协议：GUI-MCP（Graphical User Interface - Model Context Protocol），这是首个专为图形用户界面自动化而设计的 MCP 实现，兼顾标准化与隐私保护。…", "overview": "25年底，阶跃星辰升级发布了全新的AI Agent系列模型Step-GUI，包括云端模型Step-GUI、首个面向GUI Agent的MCP协议：GUI-MCP（Graphical User Interface - Model Context Protocol），这是首个专为图形用户界面自动化而设计的 MCP 实现，兼顾标准化与隐私保护。因此，我们就来解读这个MCP协议，顺便看看端侧Agent的实现架构。", "url": "https://www.cnblogs.com/rossiXYZ/p/19752402", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19752402", "series": "gui-mcp", "seriesPart": 4, "space": "deep", "coverImage": "/assets/covers/gui-mcp-04-arch.jpg", "items": [ { "title": "GUI-MCP 解读系列", "titleZh": "📚 GUI-MCP 解读系列", "summary": "共 6 篇 · 站内专栏", "category": "专栏", "seriesKey": "gui-mcp" }, { "title": "0x00 摘要", "titleZh": "0x00 摘要", "category": "章节", "url": "" }, { "title": "0x01 GUI-MCP", "titleZh": "0x01 GUI-MCP", "category": "章节", "url": "" }, { "title": "0x02 示例", "titleZh": "0x02 示例", "category": "章节", "url": "" }, { "title": "0x03 差异", "titleZh": "0x03 差异", "category": "章节", "url": "" }, { "title": "0x04 GUI-MCP 整体架构", "titleZh": "0x04 GUI-MCP 整体架构", "category": "章节", "url": "" }, { "title": "0x05 实现", "titleZh": "0x05 实现", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19752402", "external": true } ], "bodyPath": "/data/bodies/gui-mcp-04-arch.json" }, { "id": "gui-mcp-03-exec", "title": "【GUI-Agent】阶跃星辰 GUI-MCP 解读---(3)---执行层", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-03-20", "likes": 0, "featured": false, "summary": "25年底，阶跃星辰升级发布了全新的AI Agent系列模型Step-GUI，包括云端模型Step-GUI、首个面向GUI Agent的MCP协议：GUI-MCP（Graphical User Interface - Model Context Protocol），这是首个专为图形用户界面自动化而设计的 MCP 实现，兼顾标准化与隐私保护。…", "overview": "25年底，阶跃星辰升级发布了全新的AI Agent系列模型Step-GUI，包括云端模型Step-GUI、首个面向GUI Agent的MCP协议：GUI-MCP（Graphical User Interface - Model Context Protocol），这是首个专为图形用户界面自动化而设计的 MCP 实现，兼顾标准化与隐私保护。因此，我们就来解读这个MCP协议，顺便看看端侧Agent的实现架构。", "url": "https://www.cnblogs.com/rossiXYZ/p/19715713", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19715713", "series": "gui-mcp", "seriesPart": 3, "space": "deep", "coverImage": "/assets/covers/gui-mcp-03-exec.jpg", "items": [ { "title": "GUI-MCP 解读系列", "titleZh": "📚 GUI-MCP 解读系列", "summary": "共 6 篇 · 站内专栏", "category": "专栏", "seriesKey": "gui-mcp" }, { "title": "0x00 摘要", "titleZh": "0x00 摘要", "category": "章节", "url": "" }, { "title": "0x01 执行流程", "titleZh": "0x01 执行流程", "category": "章节", "url": "" }, { "title": "0x02 执行层", "titleZh": "0x02 执行层", "category": "章节", "url": "" }, { "title": "0x03 跨平台实现与统一适配层", "titleZh": "0x03 跨平台实现与统一适配层", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19715713", "external": true } ], "bodyPath": "/data/bodies/gui-mcp-03-exec.json" }, { "id": "gui-mcp-02-decision", "title": "【GUI-Agent】阶跃星辰 GUI-MCP 解读---(2)---决策层", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-03-18", "likes": 0, "featured": false, "summary": "25年底，阶跃星辰升级发布了全新的AI Agent系列模型Step-GUI，包括云端模型Step-GUI、首个面向GUI Agent的MCP协议：GUI-MCP（Graphical User Interface - Model Context Protocol），这是首个专为图形用户界面自动化而设计的 MCP 实现，兼顾标准化与隐私保护。…", "overview": "25年底，阶跃星辰升级发布了全新的AI Agent系列模型Step-GUI，包括云端模型Step-GUI、首个面向GUI Agent的MCP协议：GUI-MCP（Graphical User Interface - Model Context Protocol），这是首个专为图形用户界面自动化而设计的 MCP 实现，兼顾标准化与隐私保护。因此，我们就来解读这个MCP协议，顺便看看端侧Agent的实现架构。", "url": "https://www.cnblogs.com/rossiXYZ/p/19715693", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19715693", "series": "gui-mcp", "seriesPart": 2, "space": "deep", "coverImage": "/assets/covers/gui-mcp-02-decision.jpg", "items": [ { "title": "GUI-MCP 解读系列", "titleZh": "📚 GUI-MCP 解读系列", "summary": "共 6 篇 · 站内专栏", "category": "专栏", "seriesKey": "gui-mcp" }, { "title": "0x00 摘要", "titleZh": "0x00 摘要", "category": "章节", "url": "" }, { "title": "0x01 LocalServer", "titleZh": "0x01 LocalServer", "category": "章节", "url": "" }, { "title": "0x02 Parser0920Summary 的功能", "titleZh": "0x02 Parser0920Summary 的功能", "category": "章节", "url": "" }, { "title": "0x03 交互", "titleZh": "0x03 交互", "category": "章节", "url": "" }, { "title": "0x04 模型分发", "titleZh": "0x04 模型分发", "category": "章节", "url": "" }, { "title": "0x05 过渡到执行层", "titleZh": "0x05 过渡到执行层", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19715693", "external": true } ], "bodyPath": "/data/bodies/gui-mcp-02-decision.json" }, { "id": "gui-mcp-01-paper", "title": "【GUI-Agent】阶跃星辰 GUI-MCP 解读---(1)---论文", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-03-16", "likes": 0, "featured": true, "summary": "25年底，阶跃星辰升级发布了全新的AI Agent系列模型Step-GUI，包括云端模型Step-GUI、首个面向GUI Agent的MCP协议：GUI-MCP（Graphical User Interface - Model Context Protocol），这是首个专为图形用户界面自动化而设计的 MCP 实现，兼顾标准化与隐私保护。", "overview": "25年底，阶跃星辰升级发布了全新的AI Agent系列模型Step-GUI，包括云端模型Step-GUI、首个面向GUI Agent的MCP协议：GUI-MCP（Graphical User Interface - Model Context Protocol），这是首个专为图形用户界面自动化而设计的 MCP 实现，兼顾标准化与隐私保护。", "url": "https://www.cnblogs.com/rossiXYZ/p/19715590", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19715590", "series": "gui-mcp", "seriesPart": 1, "space": "deep", "coverImage": "/assets/covers/gui-mcp-01-paper.jpg", "items": [ { "title": "GUI-MCP 解读系列", "titleZh": "📚 GUI-MCP 解读系列", "summary": "共 6 篇 · 站内专栏", "category": "专栏", "seriesKey": "gui-mcp" }, { "title": "0x00 摘要", "titleZh": "0x00 摘要", "category": "章节", "url": "" }, { "title": "0x01 GUI Agent 的核心要素", "titleZh": "0x01 GUI Agent 的核心要素", "category": "章节", "url": "" }, { "title": "0x02 阶跃星辰论文解读", "titleZh": "0x02 阶跃星辰论文解读", "category": "章节", "url": "" }, { "title": "0x03 无MCP调用", "titleZh": "0x03 无MCP调用", "category": "章节", "url": "" }, { "title": "0x04 Agent在哪里？", "titleZh": "0x04 Agent在哪里？", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19715590", "external": true } ], "bodyPath": "/data/bodies/gui-mcp-01-paper.json" }, { "id": "openhands-14-microagents", "title": "AI Agent 框架探秘：拆解 OpenHands（14）--- Microagents", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-03-13", "likes": 0, "featured": false, "summary": "很多Agent系统会采用多智能体（multi-agent）架构，划分不同子模块/子Agent各司其职，由中央调度Agent统筹管理整个生命周期。这种模块化架构能将复杂任务拆解给最擅长该子任务的模块，发挥各模型所长，避免单一模型在某些任务上的弱点。", "overview": "很多Agent系统会采用多智能体（multi-agent）架构，划分不同子模块/子Agent各司其职，由中央调度Agent统筹管理整个生命周期。这种模块化架构能将复杂任务拆解给最擅长该子任务的模块，发挥各模型所长，避免单一模型在某些任务上的弱点。", "url": "https://www.cnblogs.com/rossiXYZ/p/19694116", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19694116", "series": "openhands", "seriesPart": 14, "space": "deep", "coverImage": "/assets/covers/openhands-14-microagents.jpg", "items": [ { "title": "OpenHands 拆解系列", "titleZh": "📚 OpenHands 拆解系列", "summary": "共 8 篇 · 站内专栏", "category": "专栏", "seriesKey": "openhands" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 需求", "titleZh": "0x01 需求", "category": "章节", "url": "" }, { "title": "0x02 基本梳理", "titleZh": "0x02 基本梳理", "category": "章节", "url": "" }, { "title": "0x03 实现", "titleZh": "0x03 实现", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19694116", "external": true } ], "bodyPath": "/data/bodies/openhands-14-microagents.json" }, { "id": "openhands-13-memory", "title": "AI Agent框架探秘：拆解 OpenHands（13）--- Memory", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-03-11", "likes": 0, "featured": false, "summary": "大模型正在从生成工具演化为具有长期交互能力的智能体，这对“记忆能力”提出了更高的要求，因为大模型的 “记忆能力”，决定了它能走多远，从单轮问答到多轮协作，从通用助手到垂直 Agent，核心都是 “能否记住关键信息、锚定核心目标”——只有记忆突破，AI才能“持续陪伴”，这是增加用户黏性的必然。", "overview": "大模型正在从生成工具演化为具有长期交互能力的智能体，这对“记忆能力”提出了更高的要求，因为大模型的 “记忆能力”，决定了它能走多远，从单轮问答到多轮协作，从通用助手到垂直 Agent，核心都是 “能否记住关键信息、锚定核心目标”——只有记忆突破，AI才能“持续陪伴”，这是增加用户黏性的必然。", "url": "https://www.cnblogs.com/rossiXYZ/p/19694082", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19694082", "series": "openhands", "seriesPart": 13, "space": "deep", "coverImage": "/assets/covers/openhands-13-memory.jpg", "items": [ { "title": "OpenHands 拆解系列", "titleZh": "📚 OpenHands 拆解系列", "summary": "共 8 篇 · 站内专栏", "category": "专栏", "seriesKey": "openhands" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 大模型 Agent 记忆系统", "titleZh": "0x01 大模型 Agent 记忆系统", "category": "章节", "url": "" }, { "title": "0x02 大模型 Agent 记忆系统分类体系", "titleZh": "0x02 大模型 Agent 记忆系统分类体系", "category": "章节", "url": "" }, { "title": "0x03 比对", "titleZh": "0x03 比对", "category": "章节", "url": "" }, { "title": "0x04 OpenHands Memory功能解析", "titleZh": "0x04 OpenHands Memory功能解析", "category": "章节", "url": "" }, { "title": "0x03 关键组件", "titleZh": "0x03 关键组件", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19694082", "external": true } ], "bodyPath": "/data/bodies/openhands-13-memory.json" }, { "id": "openhands-12-function-call", "title": "AI Agent框架探秘：拆解 OpenHands（12）--- Function call", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-03-09", "likes": 0, "featured": false, "summary": "大模型本质上是一个文本生成器，它不能直接操作系统、调用 API、访问数据库。所有这些能力都需要额外的工程实现。Agent 工具使用模式是突破大语言模型（LLM）固有局限、实现 Agent 与现实世界交互的核心架构范式，其本质是让 LLM 从单纯的文本生成器转变为具备感知、推理和行动能力的智能体，核心依托 ReAct 循环中模型对工具调用时…", "overview": "大模型本质上是一个文本生成器，它不能直接操作系统、调用 API、访问数据库。所有这些能力都需要额外的工程实现。Agent 工具使用模式是突破大语言模型（LLM）固有局限、实现 Agent 与现实世界交互的核心架构范式，其本质是让 LLM 从单纯的文本生成器转变为具备感知、推理和行动能力的智能体，核心依托 ReAct 循环中模型对工具调用时机的自主决策能力。", "url": "https://www.cnblogs.com/rossiXYZ/p/19684479", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19684479", "series": "openhands", "seriesPart": 12, "space": "deep", "coverImage": "/assets/covers/openhands-12-function-call.jpg", "items": [ { "title": "OpenHands 拆解系列", "titleZh": "📚 OpenHands 拆解系列", "summary": "共 8 篇 · 站内专栏", "category": "专栏", "seriesKey": "openhands" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 工具系统设计", "titleZh": "0x01 工具系统设计", "category": "章节", "url": "" }, { "title": "0x02 OpenHands 的设计", "titleZh": "0x02 OpenHands 的设计", "category": "章节", "url": "" }, { "title": "0x03 功能解析", "titleZh": "0x03 功能解析", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19684479", "external": true } ], "bodyPath": "/data/bodies/openhands-12-function-call.json" }, { "id": "openhands-11-runtime-components", "title": "AI Agent框架探秘：拆解 OpenHands（11）--- Runtime主要组件", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-03-05", "likes": 0, "featured": false, "summary": "因此，业界大多采用微服务架构或插件化设计，模块间通过标准化接口通信，新增功能只需开发插件并注册。", "overview": "· ActionExecutor：在 Runtime 中执行动作的核心组件\n· ActionExecutor 初始化时会根据配置加载指定的插件。插件注册到 ActionExecutor 的插件字典。\n· 当接收到动作请求时，ActionExecutor 会调用相应的方法执行动作。\n· 对于浏览动作，ActionExecutor 会使用 BrowserEnv 来处理。\n· 如果涉及插件，ActionExecutor 会通过插件系统处理\n· AgentSkillsPlugin：提供智能体技能功能的插件\n· AgentSkillsPlugin 是一个插件，继承自 Plugin 基类。\n· Runtime 初始化时，插件会被加载到插件字典中。插件通过 PluginRequirement 机制被注册到系统中。\n· 特定动作触发时调用相应插件功能。\n· BrowserEnv：浏览器环境封装，使用 BrowserGym 库。\n· ActionExecutor 在初始化时根据配置决定是否启用浏览器环境。…", "url": "https://www.cnblogs.com/rossiXYZ/p/19656848", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19656848", "series": "openhands", "seriesPart": 11, "space": "deep", "coverImage": "/assets/covers/openhands-11-runtime-components.jpg", "items": [ { "title": "OpenHands 拆解系列", "titleZh": "📚 OpenHands 拆解系列", "summary": "共 8 篇 · 站内专栏", "category": "专栏", "seriesKey": "openhands" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 三大组件", "titleZh": "0x01 三大组件", "category": "章节", "url": "" }, { "title": "0x02 数据流", "titleZh": "0x02 数据流", "category": "章节", "url": "" }, { "title": "0x03 插件系统", "titleZh": "0x03 插件系统", "category": "章节", "url": "" }, { "title": "0x04 执行系统", "titleZh": "0x04 执行系统", "category": "章节", "url": "" }, { "title": "0x05 环境", "titleZh": "0x05 环境", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19656848", "external": true } ], "bodyPath": "/data/bodies/openhands-11-runtime-components.json" }, { "id": "openhands-10-runtime", "title": "AI Agent框架探秘：拆解 OpenHands（10）--- Runtime", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-03-02", "likes": 0, "featured": false, "summary": "在 OpenHands 里，真正让“AI 想法”落地就是Runtime。它像一座可移动的实验室：四面墙把主机世界隔开，却给 Agent 留下齐全的操作台（文件、终端、网络）。Agent 只需递上一张写满指令的纸条，Runtime 便接管全部脏活：启动进程、捕捉回显、隔离风险、回收资源，再把结果封装成“观察报告”送回前台。由于所有动作都在统…", "overview": "在 OpenHands 里，真正让“AI 想法”落地就是Runtime。它像一座可移动的实验室：四面墙把主机世界隔开，却给 Agent 留下齐全的操作台（文件、终端、网络）。Agent 只需递上一张写满指令的纸条，Runtime 便接管全部脏活：启动进程、捕捉回显、隔离风险、回收资源，再把结果封装成“观察报告”送回前台。由于所有动作都在统一沙箱里完成，外部系统既不用担心权限越界，也不必为环境差异头疼。", "url": "https://www.cnblogs.com/rossiXYZ/p/19656834", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19656834", "series": "openhands", "seriesPart": 10, "space": "deep", "coverImage": "/assets/covers/openhands-10-runtime.jpg", "items": [ { "title": "OpenHands 拆解系列", "titleZh": "📚 OpenHands 拆解系列", "summary": "共 8 篇 · 站内专栏", "category": "专栏", "seriesKey": "openhands" }, { "title": "0x00 摘要", "titleZh": "0x00 摘要", "category": "章节", "url": "" }, { "title": "0x01 工作机制", "titleZh": "0x01 工作机制", "category": "章节", "url": "" }, { "title": "0x02 核心逻辑", "titleZh": "0x02 核心逻辑", "category": "章节", "url": "" }, { "title": "0x03 代码", "titleZh": "0x03 代码", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19656834", "external": true } ], "bodyPath": "/data/bodies/openhands-10-runtime.json" }, { "id": "openhands-09-agentcontroller", "title": "AI Agent框架探秘：拆解 OpenHands（9）--- AgentController", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-02-27", "likes": 0, "featured": false, "summary": "一个成熟的 Agent 系统，必须在不干扰 Agent 自主决策的前提下，提供外部管控接口。比如用户可能需要暂停任务进行参数调整，或是在发现错误时终止执行，甚至在任务中途切换代理角色。这种全生命周期的可控性，需要工作流层设计出灵活的状态机和事件触发机制，在自主性与可控性之间找到完美平衡。", "overview": "一个成熟的 Agent 系统，必须在不干扰 Agent 自主决策的前提下，提供外部管控接口。比如用户可能需要暂停任务进行参数调整，或是在发现错误时终止执行，甚至在任务中途切换代理角色。这种全生命周期的可控性，需要工作流层设计出灵活的状态机和事件触发机制，在自主性与可控性之间找到完美平衡。", "url": "https://www.cnblogs.com/rossiXYZ/p/19639111", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19639111", "series": "openhands", "seriesPart": 9, "space": "deep", "coverImage": "/assets/covers/openhands-09-agentcontroller.jpg", "items": [ { "title": "OpenHands 拆解系列", "titleZh": "📚 OpenHands 拆解系列", "summary": "共 8 篇 · 站内专栏", "category": "专栏", "seriesKey": "openhands" }, { "title": "0x00 概要", "titleZh": "0x00 概要", "category": "章节", "url": "" }, { "title": "0x01 为何需要 AgentController ?", "titleZh": "0x01 为何需要 AgentController ?", "category": "章节", "url": "" }, { "title": "0x02 AgentController", "titleZh": "0x02 AgentController", "category": "章节", "url": "" }, { "title": "0x03 重点功能", "titleZh": "0x03 重点功能", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19639111", "external": true } ], "bodyPath": "/data/bodies/openhands-09-agentcontroller.json" }, { "id": "openhands-08-codeact", "title": "AI Agent框架探秘：拆解 OpenHands（8）--- CodeActAgent", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-02-25", "likes": 0, "featured": false, "summary": "在 OpenHands 智能框架的生态中，CodeActAgent 占据着核心地位，它是基于 CodeAct 理念构建的核心代理模块。其设计初衷极具巧思：将各类复杂任务统一转化为 “代码执行” 的形式来完成，同时兼顾自然语言对话的交互特性。这一设计既保障了任务执行的精准性与高效性，又为人类与智能代理的协作提供了灵活空间，使其成为框架中处理…", "overview": "在 OpenHands 智能框架的生态中，CodeActAgent 占据着核心地位，它是基于 CodeAct 理念构建的核心代理模块。其设计初衷极具巧思：将各类复杂任务统一转化为 “代码执行” 的形式来完成，同时兼顾自然语言对话的交互特性。这一设计既保障了任务执行的精准性与高效性，又为人类与智能代理的协作提供了灵活空间，使其成为框架中处理自动化编程、数据处理等复杂场景的核心载体。", "url": "https://www.cnblogs.com/rossiXYZ/p/19620422", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19620422", "series": "openhands", "seriesPart": 8, "space": "deep", "coverImage": "/assets/covers/openhands-08-codeact.jpg", "items": [ { "title": "OpenHands 拆解系列", "titleZh": "📚 OpenHands 拆解系列", "summary": "共 8 篇 · 站内专栏", "category": "专栏", "seriesKey": "openhands" }, { "title": "0x00 摘要", "titleZh": "0x00 摘要", "category": "章节", "url": "" }, { "title": "0x01 背景", "titleZh": "0x01 背景", "category": "章节", "url": "" }, { "title": "0x02 定义", "titleZh": "0x02 定义", "category": "章节", "url": "" }, { "title": "0x03 工作流程", "titleZh": "0x03 工作流程", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19620422", "external": true } ], "bodyPath": "/data/bodies/openhands-08-codeact.json" }, { "id": "openhands-core", "title": "AI Agent框架探秘：拆解 OpenHands（1）——核心理念", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "date": "2026-01-19", "likes": 0, "featured": true, "summary": "写这个系列的原因，是因为和几个同学讨论Agent系统，但是总觉得自己讲得浮于表面，兄弟们好像也没有很理解。因此决定再系统梳理一下。", "overview": "写这个系列的原因，是因为和几个同学讨论Agent系统，但是总觉得自己讲得浮于表面，兄弟们好像也没有很理解。因此决定再系统梳理一下。", "url": "https://www.cnblogs.com/rossiXYZ/p/19497161", "source": "cnblogs", "sourceLabel": "博客园 · 罗西的思考", "sourcePostId": "19497161", "series": "openhands", "seriesPart": 1, "space": "deep", "coverImage": "/assets/covers/openhands-core.jpg", "items": [ { "title": "OpenHands 拆解系列", "titleZh": "📚 OpenHands 拆解系列", "summary": "共 8 篇 · 站内专栏", "category": "专栏", "seriesKey": "openhands" }, { "title": "0x00 摘要", "titleZh": "0x00 摘要", "category": "章节", "url": "" }, { "title": "0x01 背景", "titleZh": "0x01 背景", "category": "章节", "url": "" }, { "title": "0x02 AI Agent 系统", "titleZh": "0x02 AI Agent 系统", "category": "章节", "url": "" }, { "title": "0x03 OpenHands 架构概念图谱", "titleZh": "0x03 OpenHands 架构概念图谱", "category": "章节", "url": "" }, { "title": "0xFF 参考", "titleZh": "0xFF 参考", "category": "章节", "url": "" }, { "title": "原文出处（博客园）", "titleZh": "原文出处（博客园）", "summary": "博客园 · 罗西的思考", "category": "外链", "url": "https://www.cnblogs.com/rossiXYZ/p/19497161", "external": true } ], "bodyPath": "/data/bodies/openhands-core.json" }, { "id": "writing-after-ai", "title": "AI 写得比你好之后，作家还剩下什么？Alain de Botton 的回答", "category": "talk", "categoryName": "大佬讲话", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "likes": 0, "summary": "当机器能快速生成流畅文字，人的价值更集中在经验、判断与真正值得表达的问题上。", "coverImage": "/assets/covers/writing-after-ai.jpg", "items": [], "overview": "当机器能快速生成流畅文字，人的价值更集中在经验、判断与真正值得表达的问题上。" }, { "id": "how-to-speak", "title": "MIT 教授连续 40 年最火的一堂课：不教编程，教你怎么说话", "category": "talk", "categoryName": "大佬讲话", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "likes": 0, "summary": "从开场、节奏、结构和听众注意力出发，拆解如何把复杂想法讲得清楚且可信。", "coverImage": "/assets/covers/how-to-speak.jpg", "items": [], "overview": "从开场、节奏、结构和听众注意力出发，拆解如何把复杂想法讲得清楚且可信。" }, { "id": "tao-math-ai", "title": "陶哲轩：AI 做数学成功率只有 2%，但这可能是个好消息", "category": "paper", "categoryName": "AI 论文", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "likes": 0, "summary": "低成功率不等于没有价值；在高难度数学探索中，可验证的小概率突破仍可能产生研究增量。", "coverImage": "/assets/covers/tao-math-ai.jpg", "items": [], "overview": "低成功率不等于没有价值；在高难度数学探索中，可验证的小概率突破仍可能产生研究增量。" }, { "id": "tools-for-thought", "title": "Notion 联合创始人 Ivan Zhao：蒸汽、钢铁与无限思维", "category": "talk", "categoryName": "大佬讲话", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "likes": 1, "summary": "回顾工具如何放大人的体力与认知，并讨论软件在 AI 时代可能成为怎样的思维基础设施。", "coverImage": "/assets/covers/tools-for-thought.jpg", "items": [], "overview": "回顾工具如何放大人的体力与认知，并讨论软件在 AI 时代可能成为怎样的思维基础设施。" }, { "id": "memory-system", "title": "我逆向工程了 ChatGPT 的记忆系统", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "likes": 3, "summary": "从可观察行为推测记忆的写入、召回和更新机制，并分析长期上下文系统的设计取舍。", "coverImage": "/assets/covers/memory-system.jpg", "items": [], "overview": "从可观察行为推测记忆的写入、召回和更新机制，并分析长期上下文系统的设计取舍。" }, { "id": "beyond-scaling", "title": "Ilya：Scaling 时代结束，下一阶段如何让模型学得像人", "category": "trend", "categoryName": "未来趋势", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "likes": 0, "summary": "单纯扩大训练规模的边际收益下降后，研究重点可能转向更有效的学习机制与推理过程。", "coverImage": "/assets/covers/beyond-scaling.jpg", "items": [], "overview": "单纯扩大训练规模的边际收益下降后，研究重点可能转向更有效的学习机制与推理过程。" }, { "id": "agent-arch", "title": "Google AI Agent 架构白皮书：设计、部署和治理下一代 Agent", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "likes": 1, "summary": "梳理代理系统的模型、工具、记忆、编排、观测和治理层，为工程落地提供结构化视图。", "coverImage": "/assets/covers/agent-arch.jpg", "items": [], "overview": "梳理代理系统的模型、工具、记忆、编排、观测和治理层，为工程落地提供结构化视图。" }, { "id": "mckinsey-ai", "title": "麦肯锡报告：AI 现状进入智能体时代的临界点", "category": "research", "categoryName": "研究报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "likes": 1, "summary": "从企业采用、价值实现和组织准备度观察生成式 AI 与智能体的扩散阶段。", "coverImage": "/assets/covers/mckinsey-ai.jpg", "items": [], "overview": "从企业采用、价值实现和组织准备度观察生成式 AI 与智能体的扩散阶段。" }, { "id": "ai-learning", "title": "Google《AI 与学习的未来》：人工智能如何影响教育", "category": "research", "categoryName": "研究报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "likes": 0, "summary": "讨论个性化学习、教师角色、评价机制和教育公平，并强调需要可验证的教学效果。", "coverImage": "/assets/covers/ai-learning.jpg", "items": [], "overview": "讨论个性化学习、教师角色、评价机制和教育公平，并强调需要可验证的教学效果。" }, { "id": "context-eng", "title": "让大模型理解世界的工程学：Context Engineering 指南", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "likes": 3, "summary": "把提示词扩展为包含检索、工具、记忆和状态管理的完整上下文系统。", "coverImage": "/assets/covers/context-eng.jpg", "items": [], "overview": "把提示词扩展为包含检索、工具、记忆和状态管理的完整上下文系统。" }, { "id": "atlas-browser", "title": "OpenAI 揭秘 AI 浏览器 ChatGPT Atlas 的构建方式", "category": "technical", "categoryName": "技术报告", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "likes": 0, "summary": "聚焦浏览器环境中的页面理解、操作规划、权限边界与用户控制等工程问题。", "coverImage": "/assets/covers/atlas-browser.jpg", "items": [], "overview": "聚焦浏览器环境中的页面理解、操作规划、权限边界与用户控制等工程问题。" }, { "id": "synthetic-research", "title": "能否用 AI 代替人类做产品购买意愿调查？一次低成本实验", "category": "paper", "categoryName": "AI 论文", "cover": "cover-photo", "coverText": "", "coverSub": "", "time": "深度", "likes": 1, "summary": "探索合成受访者用于早期概念筛选的可行性，同时提醒偏差校准和真实用户验证仍不可替代。", "coverImage": "/assets/covers/synthetic-research.jpg", "items": [], "overview": "探索合成受访者用于早期概念筛选的可行性，同时提醒偏差校准和真实用户验证仍不可替代。" } ], "stats": { "frontier": 14, "toolbox": 12, "deep": 72, "series": 11, "dailyDate": "2026-07-15" } }