[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-77506":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":9,"language":10,"languages":9,"totalLinesOfCode":9,"stars":11,"forks":12,"watchers":13,"openIssues":14,"contributorsCount":14,"subscribersCount":14,"size":14,"stars1d":15,"stars7d":16,"stars30d":17,"stars90d":14,"forks30d":14,"starsTrendScore":18,"compositeScore":19,"rankGlobal":9,"rankLanguage":9,"license":20,"archived":21,"fork":21,"defaultBranch":22,"hasWiki":23,"hasPages":21,"topics":24,"createdAt":9,"pushedAt":9,"updatedAt":25,"readmeContent":26,"aiSummary":27,"trendingCount":14,"starSnapshotCount":14,"syncStatus":28,"lastSyncTime":29,"discoverSource":30},77506,"Agent-Learning-Hub","datawhalechina\u002FAgent-Learning-Hub","datawhalechina","AI Agent 学习路线与资料库收集",null,"HTML",3490,354,7,0,83,800,3378,433,109.65,"MIT License",false,"main",true,[],"2026-06-12 04:01:21","# Agent Learning Hub\n\nA curated AI Agent learning roadmap for people who want to build useful, reliable agents instead of collecting random links.\n\n这个仓库只维护一个核心展示面：README。目标是把社区里优秀分享、官方博客、论文、开源项目和真实工程经验，整理成一份可以照着执行的 AI Agent 学习 todo list。\n\n## Maintainer\n\nCurated by [陈思州](https:\u002F\u002Fgithub.com\u002Fjjyaoao) (Datawhale 成员) \u003Ca href=\"https:\u002F\u002Fwww.xiaohongshu.com\u002Fuser\u002Fprofile\u002F67b9cc34000000000e013517\" target=\"_blank\">\u003Cimg alt=\"Static Badge\" src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FRednote-小红书-e93c49\">\u003C\u002Fa>\n\n## How To Use\n\n- 如果你是新手：按「Learning Todo List」从上到下做，每完成一项就打勾。\n- 如果你已经会 LLM 应用：从 Stage 2 或 Stage 3 开始，重点补 Agent loop、工具调用、评测和工程化。\n- 如果你想做项目：直接看「Project Ladder」，每一档做一个可运行作品。\n- 如果你只想找资料：看「Curated Resources」，优先读官方文档和经典论文。\n\n## What To Learn Now\n\nAgent 领域变化很快。当前更值得投入的不是老式“角色扮演多 agent 框架”，而是这些更贴近真实生产力的方向：\n\n| Priority | Learn | Why |\n| --- | --- | --- |\n| 1 | Claude Code \u002F Codex-style coding agents | 真实代码库、shell、文件编辑、测试、权限、上下文压缩，是最好的 agent 工程样本。 |\n| 2 | Agent harness engineering | agent 的能力很大一部分来自 harness：工具协议、权限、状态、反馈、回放、CI、评测。 |\n| 3 | OpenClaw \u002F Hermes-style personal agents | 长运行、本地优先、跨应用、记忆、skills、消息入口，更像“个人操作系统”。 |\n| 4 | Skills \u002F MCP \u002F A2A \u002F ACP | skills 负责能力复用，MCP 连接工具，A2A 连接 agent，ACP 连接宿主应用。 |\n| 5 | Evaluation and safety | 没有 eval、trace、权限边界的 agent 只能算 demo。 |\n\n不建议把精力重押在已经泛化成模板的老式 crew\u002Frole-play 框架上。它们可以了解，但不应成为主线。\n\n## Learning Todo List\n\n### Stage 0: Understand What An Agent Is\n\n- [ ] 区分 chatbot、workflow、agent、multi-agent。\n- [ ] 理解 agent 的基本循环：observe -> think -> act -> observe。\n- [ ] 明白什么时候不该用 agent：任务可预测、流程稳定、普通脚本能解决时，agent 反而增加不确定性。\n- [ ] 读完 [Anthropic: Building effective agents](https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Fbuilding-effective-agents)。\n- [ ] 读完 [OpenAI: A practical guide to building agents](https:\u002F\u002Fopenai.com\u002Fbusiness\u002Fguides-and-resources\u002Fa-practical-guide-to-building-ai-agents\u002F)。\n\n产出：写一页短笔记，回答「我的场景为什么需要 agent，而不是普通 workflow？」\n\n### Stage 1: Build A Minimal Agent Loop\n\n- [ ] 会用一个 LLM API 完成普通对话。\n- [ ] 会让模型输出结构化 JSON。\n- [ ] 会定义一个工具函数，例如 search、calculator、read_file。\n- [ ] 会解析模型的 tool call \u002F function call。\n- [ ] 会执行工具，并把工具结果喂回模型。\n- [ ] 会给 agent loop 加最大步数、超时和错误处理。\n\n推荐阅读：\n\n- [OpenAI Function Calling](https:\u002F\u002Fplatform.openai.com\u002Fdocs\u002Fguides\u002Ffunction-calling)\n- [Gemini API Function Calling](https:\u002F\u002Fai.google.dev\u002Fgemini-api\u002Fdocs\u002Ffunction-calling)\n- [Claude Tool Use](https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fagents-and-tools\u002Ftool-use\u002Foverview)\n\n产出：一个 50-150 行的最小 agent，可以选择工具、执行工具、返回最终答案。\n\n### Stage 2: Learn Tool Use, RAG, And Memory\n\n- [ ] 会做检索增强生成：chunk、embed、retrieve、answer with citations。\n- [ ] 会把搜索、数据库、文件、浏览器、代码执行接成工具。\n- [ ] 会区分短期上下文、会话记忆、长期记忆。\n- [ ] 会处理工具失败、空结果、重复调用、幻觉引用。\n- [ ] 会让 agent 在回答里给出来源或证据。\n\n推荐阅读：\n\n- [LlamaIndex Agents](https:\u002F\u002Fdocs.llamaindex.ai\u002Fen\u002Fstable\u002Fuse_cases\u002Fagents\u002F)\n- [LangChain Docs](https:\u002F\u002Fdocs.langchain.com\u002F)\n- [Gemini API Code Execution](https:\u002F\u002Fai.google.dev\u002Fgemini-api\u002Fdocs\u002Fcode-execution)\n- [Model Context Protocol](https:\u002F\u002Fmodelcontextprotocol.io\u002F)\n\n开源项目参考：\n\n| Project | Why It Fits Stage 2 |\n| --- | --- |\n| [GPT Researcher](https:\u002F\u002Fgithub.com\u002Fassafelovic\u002Fgpt-researcher) | 最接近“资料研究助手”的成品：搜索、抓取、筛选、引用、生成长报告。 |\n| [Open Deep Research](https:\u002F\u002Fgithub.com\u002Flangchain-ai\u002Fopen_deep_research) | LangGraph 写的 deep research 示例，适合学多轮搜索、状态管理和引用输出。 |\n| [STORM](https:\u002F\u002Fgithub.com\u002Fstanford-oval\u002Fstorm) | Stanford OVAL 的研究写作系统，适合学 outline、question asking、多视角资料综合。 |\n| [Khoj](https:\u002F\u002Fgithub.com\u002Fkhoj-ai\u002Fkhoj) | 个人 second brain，适合学本地文档、网页、语义搜索和长期记忆。 |\n| [Onyx](https:\u002F\u002Fgithub.com\u002Fonyx-dot-app\u002Fonyx) | 企业级 RAG\u002Fsearch assistant，适合学 connectors、hybrid search、权限和生产化。 |\n| [AnythingLLM](https:\u002F\u002Fgithub.com\u002FMintplex-Labs\u002Fanything-llm) | 本地 RAG + agents 产品，适合初学者快速理解完整应用形态。 |\n| [RAGFlow](https:\u002F\u002Fgithub.com\u002Finfiniflow\u002Fragflow) | 文档理解型 RAG 引擎，适合学 ingestion、chunking、retrieval、grounded answer。 |\n| [mem0](https:\u002F\u002Fgithub.com\u002Fmem0ai\u002Fmem0) | 记忆层组件，适合学如何给 agent 加长期 memory。 |\n| [Letta](https:\u002F\u002Fgithub.com\u002Fletta-ai\u002Fletta) | 面向 stateful agents 的 memory\u002Fcontext 平台，适合学上下文管理。 |\n\n产出：一个资料研究助手，输入主题后自动搜索、筛选、总结并输出引用链接。\n\n### Stage 3: Study One Modern Agent Harness\n\n先选一个现代 agent 系统学深。这里的重点不是“框架 API 怎么调”，而是它如何组织工具、上下文、权限、状态、日志、子任务和反馈。\n\n| System | Best For | Learn This If You Want To |\n| --- | --- | --- |\n| [Claude Code Docs](https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Foverview) | Coding agent product | 学真实 coding agent 的 CLI、工具、权限、hooks、subagents、MCP。 |\n| [learn-claude-code](https:\u002F\u002Fgithub.com\u002FshareAI-lab\u002Flearn-claude-code) | From-scratch agent harness | 从 0 到 1 复刻 Claude Code-like harness。 |\n| [claw0](https:\u002F\u002Fgithub.com\u002FshareAI-lab\u002Fclaw0) | From-scratch OpenClaw gateway | 从 agent loop 一路构建 session、channel、gateway、memory、heartbeat、delivery、resilience、concurrency。 |\n| [hello-agents](https:\u002F\u002Fgithub.com\u002Fdatawhalechina\u002Fhello-agents) | Chinese agent tutorial | 从零开始构建智能体，适合系统补 Agent 原理与实践。 |\n| [OpenClaw](https:\u002F\u002Fgithub.com\u002Fopenclaw\u002Fopenclaw) | Local-first personal agent | 学本地长运行 agent、skills、消息入口、系统工具和安全边界。 |\n| [Hermes Agent](https:\u002F\u002Fgithub.com\u002FNousResearch\u002Fhermes-agent) | Self-hosted growing agent | 学长期记忆、skills、toolsets、多平台消息网关和迁移能力。 |\n| [CyberClaw](https:\u002F\u002Fgithub.com\u002Fttguy0707\u002FCyberClaw) | Transparent agent architecture | 学全行为审计、两段式安全调用、双水位记忆和心跳任务。 |\n| [LangGraph](https:\u002F\u002Flangchain-ai.github.io\u002Flanggraph\u002F) | Stateful graph orchestration | 学状态图、可恢复执行和可控编排。 |\n\n- [ ] 读懂一个 agent harness 的目录结构。\n- [ ] 找出它的 agent loop、tool registry、permission gate、session store、context compaction。\n- [ ] 跑通它的最小示例，并加一个你自己的工具。\n- [ ] 观察一次完整 trace，解释每一步为什么发生。\n- [ ] 把同一个任务分别用「裸 agent loop」和「harness」实现，对比差异。\n\n产出：一个可调试的 agent harness demo，包含 README、运行步骤、示例输入输出和失败记录。\n\n### Stage 4: Multi-Agent Is Coordination, Not Magic\n\n- [ ] 理解 planner \u002F executor \u002F reviewer \u002F critic \u002F router 等常见角色。\n- [ ] 学会用 supervisor 或 graph 管理多 agent，而不是让 agent 随便聊天。\n- [ ] 会定义每个 agent 的职责边界、输入输出 schema、停止条件。\n- [ ] 会处理循环、争论、任务漂移、上下文膨胀。\n- [ ] 会判断什么时候单 agent 更好。\n\n推荐阅读：\n\n- [Claude Code Subagents](https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Fsub-agents)\n- [Claude Code Hooks](https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Fhooks)\n- [Google Agent Development Kit](https:\u002F\u002Fgoogle.github.io\u002Fadk-docs\u002F)\n- [Agent2Agent Protocol](https:\u002F\u002Fgoogle-a2a.github.io\u002FA2A\u002Fspecification\u002F)\n- [Agent Client Protocol](https:\u002F\u002Fagentclientprotocol.com\u002F)\n\n产出：一个小型多 agent 系统，例如 research -> write -> review -> revise。\n\n### Stage 5: Learn Skills, Protocols, And Capability Packaging\n\n现代 agent 的能力不只来自模型和工具，也来自可复用的 skills。一个好的 skill 像一份小型操作手册：告诉 agent 什么时候使用、怎么使用、需要哪些脚本\u002F资源、如何验证结果。\n\n- [ ] 理解 Skill 和 Tool 的区别：tool 是可调用接口，skill 是可复用流程知识。\n- [ ] 理解 Skill 和 Prompt 的区别：prompt 通常是一次性指令，skill 是可发现、可版本化、可分发的能力包。\n- [ ] 理解 Skill 和 MCP 的区别：MCP 接入外部工具\u002F数据源，skill 告诉 agent 如何完成一类任务。\n- [ ] 阅读 Claude Code Skills 的文件结构和触发机制。\n- [ ] 阅读 OpenClaw Skills 的加载、作用域和安全边界。\n- [ ] 写一个最小 `SKILL.md`，包含 name、description、何时使用、步骤、验收标准。\n- [ ] 给 skill 加一个脚本或模板文件，并说明 agent 什么时候才需要加载它。\n- [ ] 给 skill 写一个 smoke test，验证它是否真的提升任务成功率。\n\n推荐阅读：\n\n- [Claude Code Skills](https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Fskills)\n- [Claude Agent Skills](https:\u002F\u002Fdocs.claude.com\u002Fen\u002Fdocs\u002Fagents-and-tools\u002Fagent-skills)\n- [Claude Code Agent SDK Skills](https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Fagent-sdk\u002Fskills)\n- [OpenClaw Skills](https:\u002F\u002Fgithub.com\u002Fopenclaw\u002Fopenclaw\u002Fblob\u002Fmain\u002Fdocs\u002Ftools\u002Fskills.md)\n- [Model Context Protocol](https:\u002F\u002Fmodelcontextprotocol.io\u002F)\n- [Agent2Agent Protocol](https:\u002F\u002Fgoogle-a2a.github.io\u002FA2A\u002Fspecification\u002F)\n- [Agent Client Protocol](https:\u002F\u002Fagentclientprotocol.com\u002F)\n\n产出：一个可复用 skill，例如 code-review、research-report、migration-helper、pdf-extraction 或 release-note-writer。\n\n### Stage 6: Browser And Computer-Use Agents\n\n- [ ] 理解 browser agent 和普通 API tool 的区别。\n- [ ] 会用 Playwright 或 browser-use 做网页观察和点击。\n- [ ] 会给浏览器操作加安全限制：不登录敏感账号、不越权、不绕过平台规则。\n- [ ] 会处理页面变化、弹窗、加载失败、元素定位失败。\n- [ ] 会记录截图、DOM、动作日志，方便复盘。\n\n推荐阅读：\n\n- [Claude Computer Use](https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fagents-and-tools\u002Ftool-use\u002Fcomputer-use-tool)\n- [browser-use](https:\u002F\u002Fgithub.com\u002Fbrowser-use\u002Fbrowser-use)\n- [WebArena](https:\u002F\u002Farxiv.org\u002Fabs\u002F2307.13854)\n- [VisualWebArena](https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.13649)\n\n产出：一个只操作公开网页的 browser agent，例如打开网页、提取信息、生成摘要。\n\n### Stage 7: Evaluation, Observability, And Safety\n\n- [ ] 为 agent 准备固定测试集，而不是只看 demo。\n- [ ] 记录成功率、失败原因、工具调用次数、成本、延迟。\n- [ ] 会看 trace，知道失败发生在 prompt、工具、检索、模型还是状态管理。\n- [ ] 给危险工具加人工确认，例如发邮件、删文件、付款、发布内容。\n- [ ] 了解 prompt injection、data exfiltration、tool abuse 等风险。\n- [ ] 会用回归测试防止 prompt 或工具改动后能力退化。\n\n推荐阅读：\n\n- [OpenAI Evals](https:\u002F\u002Fplatform.openai.com\u002Fdocs\u002Fguides\u002Fevals)\n- [OpenAI Agent platform](https:\u002F\u002Fopenai.com\u002Fagent-platform\u002F)\n- [LangSmith](https:\u002F\u002Fdocs.smith.langchain.com\u002F)\n- [AgentBench](https:\u002F\u002Farxiv.org\u002Fabs\u002F2308.03688)\n- [SWE-bench](https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.06770)\n\n产出：一个 agent eval 表格，至少包含 20 个任务、期望结果、实际结果、失败分类。\n\n### Stage 8: Ship A Real Agent\n\n- [ ] 有明确用户、明确任务、明确成功标准。\n- [ ] 有日志、trace、错误重试、超时、成本上限。\n- [ ] 有权限边界和人工确认机制。\n- [ ] 有部署方式：CLI、Web app、Slack bot、GitHub Action 或后台任务。\n- [ ] 有 README：怎么运行、怎么配置 key、怎么扩展工具、有哪些限制。\n\n产出：一个别人能 clone 下来跑的 agent 项目。\n\n## Project Ladder\n\n| Level | Project | What You Learn |\n| --- | --- | --- |\n| 1 | Calculator Agent | 最小 tool call loop |\n| 2 | Web Research Agent | 搜索、筛选、引用、总结 |\n| 3 | PDF QA Agent | RAG、chunk、retrieval、citation |\n| 4 | Coding Review Agent | 读取 diff、风险排序、测试建议 |\n| 5 | Browser Agent | 页面观察、点击、提取、失败恢复 |\n| 6 | Claude Code-like Nano Agent | shell、文件编辑、权限、session、compact |\n| 7 | OpenClaw-like Gateway | channel、routing、session、memory、heartbeat、delivery |\n| 8 | Reusable Skill Pack | SKILL.md、脚本、模板、触发条件、smoke test |\n| 9 | Multi-Agent Writer | planner、writer、reviewer 协作 |\n| 10 | Personal Agent | OpenClaw\u002FHermes-style 记忆、skills、消息入口 |\n| 11 | Production Harness | evals、trace、权限、CI、runner、回放 |\n\n## Curated Resources\n\n### Official Guides And Blogs\n\n| Resource | Why It Matters |\n| --- | --- |\n| [Anthropic: Building effective agents](https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Fbuilding-effective-agents) | Agent 设计入门必读，讲清 workflow 和 agent 的边界。 |\n| [Claude Code Overview](https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Foverview) | 当前最值得研究的 coding agent 产品文档入口。 |\n| [Claude Code Subagents](https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Fsub-agents) | 学任务拆分、上下文隔离、专用子代理。 |\n| [Claude Code Hooks](https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Fhooks) | 学如何拦截、校验和扩展 agent 行为。 |\n| [Claude Code GitHub Actions](https:\u002F\u002Fdocs.claude.com\u002Fen\u002Fdocs\u002Fclaude-code\u002Fgithub-actions) | 学 coding agent 如何进入 PR \u002F issue 工作流。 |\n| [Claude Code Advanced Patterns](https:\u002F\u002Fresources.anthropic.com\u002Fhubfs\u002FClaude%20Code%20Advanced%20Patterns_%20Subagents%2C%20MCP%2C%20and%20Scaling%20to%20Real%20Codebases.pdf) | Anthropic 官方材料，覆盖 subagents、MCP 和真实代码库扩展。 |\n| [OpenAI: A practical guide to building agents](https:\u002F\u002Fopenai.com\u002Fbusiness\u002Fguides-and-resources\u002Fa-practical-guide-to-building-ai-agents\u002F) | 面向产品和工程团队的 agent 落地指南。 |\n| [OpenAI: New tools for building agents](https:\u002F\u002Fopenai.com\u002Findex\u002Fnew-tools-for-building-agents\u002F) | Responses API、Agents SDK、工具和 tracing 的官方介绍。 |\n| [OpenAI Agents SDK](https:\u002F\u002Fplatform.openai.com\u002Fdocs\u002Fguides\u002Fagents-sdk\u002F) | OpenAI 原生 agent 开发入口。 |\n| [Claude Tool Use](https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fagents-and-tools\u002Ftool-use\u002Foverview) | Claude 工具调用机制。 |\n| [Claude Computer Use](https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fagents-and-tools\u002Ftool-use\u002Fcomputer-use-tool) | 电脑操作类 agent 的官方参考。 |\n| [Gemini Function Calling](https:\u002F\u002Fai.google.dev\u002Fgemini-api\u002Fdocs\u002Ffunction-calling) | Gemini 工具调用官方文档。 |\n| [Gemini Code Execution](https:\u002F\u002Fai.google.dev\u002Fgemini-api\u002Fdocs\u002Fcode-execution) | Gemini 代码执行工具。 |\n| [Google Agent Development Kit](https:\u002F\u002Fgoogle.github.io\u002Fadk-docs\u002F) | Google 的 agent 开发框架。 |\n| [Model Context Protocol](https:\u002F\u002Fmodelcontextprotocol.io\u002F) | Agent 连接工具和数据源的重要协议。 |\n\n### Project Map\n\n项目不要按 star 数乱读，建议按学习目的分层：\n\n| Layer | Study These | Learn |\n| --- | --- | --- |\n| Build From Scratch | [learn-claude-code](https:\u002F\u002Fgithub.com\u002FshareAI-lab\u002Flearn-claude-code), [claw0](https:\u002F\u002Fgithub.com\u002FshareAI-lab\u002Fclaw0), [hello-agents](https:\u002F\u002Fgithub.com\u002Fdatawhalechina\u002Fhello-agents) | agent loop、tool registry、session、context compaction、gateway、trace、subagents。 |\n| Personal \u002F Always-On Agents | [OpenClaw](https:\u002F\u002Fgithub.com\u002Fopenclaw\u002Fopenclaw), [Hermes Agent](https:\u002F\u002Fgithub.com\u002FNousResearch\u002Fhermes-agent), [CyberClaw](https:\u002F\u002Fgithub.com\u002Fttguy0707\u002FCyberClaw) | 长运行、skills、记忆、消息入口、权限、安全审计。 |\n| Coding Agents | [Claude Code](https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Foverview), [OpenAI Codex](https:\u002F\u002Fgithub.com\u002Fopenai\u002Fcodex), [OpenCode](https:\u002F\u002Fgithub.com\u002Fopencode-ai\u002Fopencode), [OpenHands](https:\u002F\u002Fgithub.com\u002FAll-Hands-AI\u002FOpenHands), [SWE-agent](https:\u002F\u002Fgithub.com\u002FSWE-agent\u002FSWE-agent) | 真实代码库编辑、shell、测试、sandbox、PR 工作流。 |\n| Deep Research \u002F RAG Agents | [DeerFlow](https:\u002F\u002Fgithub.com\u002Fbytedance\u002Fdeer-flow), [LlamaIndex](https:\u002F\u002Fdocs.llamaindex.ai\u002F) | 搜索、抓取、检索、rerank、引用、报告生成。 |\n| Tutorial Encyclopedias | [GenAI_Agents](https:\u002F\u002Fgithub.com\u002FNirDiamant\u002FGenAI_Agents), [hello-agents](https:\u002F\u002Fgithub.com\u002Fdatawhalechina\u002Fhello-agents), [smolagents](https:\u002F\u002Fgithub.com\u002Fhuggingface\u002Fsmolagents), [agents-towards-production](https:\u002F\u002Fgithub.com\u002FNirDiamant\u002Fagents-towards-production) | 横向看 ReAct、Plan-and-Execute、Multi-Agent、production patterns。 |\n| Browser \u002F Multimodal Agents | [browser-use](https:\u002F\u002Fgithub.com\u002Fbrowser-use\u002Fbrowser-use), [UI-TARS-desktop](https:\u002F\u002Fgithub.com\u002Fbytedance\u002FUI-TARS-desktop) | 浏览器\u002F桌面操作、视觉理解、动作空间、失败恢复。 |\n\n### Skills, Protocols, And Tooling\n\n| Concept | Learn From | What It Solves |\n| --- | --- | --- |\n| Skills | [Claude Code Skills](https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Fskills), [OpenClaw Skills](https:\u002F\u002Fgithub.com\u002Fopenclaw\u002Fopenclaw\u002Fblob\u002Fmain\u002Fdocs\u002Ftools\u002Fskills.md) | 把一类任务的流程知识、脚本、模板和验收标准打包成可复用能力。 |\n| MCP | [Model Context Protocol](https:\u002F\u002Fmodelcontextprotocol.io\u002F) | 让 agent 标准化连接外部工具、数据源和服务。 |\n| A2A | [Agent2Agent Protocol](https:\u002F\u002Fgoogle-a2a.github.io\u002FA2A\u002Fspecification\u002F) | 让不同 agent 之间发现、通信和协作。 |\n| ACP | [Agent Client Protocol](https:\u002F\u002Fagentclientprotocol.com\u002F) | 让编辑器、终端、IDE、宿主应用和 agent 之间形成统一接口。 |\n| Skill Quality | [SWE-Skills-Bench](https:\u002F\u002Farxiv.org\u002Fabs\u002F2603.15401), [Agent Skills analysis](https:\u002F\u002Farxiv.org\u002Fabs\u002F2602.08004) | 评估 skills 是否真的提升成功率，而不是制造新的 prompt 噪声。 |\n\n### Modern Agent Systems\n\n| System | Why It Is Useful |\n| --- | --- |\n| [Claude Code](https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Foverview) | 学 coding agent 的产品化形态：shell、文件、权限、hooks、subagents、MCP。 |\n| [learn-claude-code](https:\u002F\u002Fgithub.com\u002FshareAI-lab\u002Flearn-claude-code) | 从零构建 Claude Code-like nano agent，适合理解 harness engineering。 |\n| [claw0](https:\u002F\u002Fgithub.com\u002FshareAI-lab\u002Fclaw0) | 从零构建 OpenClaw-like agent gateway，覆盖 channel、routing、session、memory、delivery、concurrency。 |\n| [hello-agents](https:\u002F\u002Fgithub.com\u002Fdatawhalechina\u002Fhello-agents) | 中文智能体系统教程，适合从零开始补齐 Agent 原理与实践。 |\n| [OpenClaw](https:\u002F\u002Fgithub.com\u002Fopenclaw\u002Fopenclaw) | 本地优先的个人 agent，适合研究长运行 agent 和系统级工具调用。 |\n| [Hermes Agent](https:\u002F\u002Fgithub.com\u002FNousResearch\u002Fhermes-agent) | 自托管、长期记忆、skills、消息网关和 toolsets。 |\n| [CyberClaw](https:\u002F\u002Fgithub.com\u002Fttguy0707\u002FCyberClaw) | 透明可控 agent 架构，适合研究审计、两段式执行和安全边界。 |\n| [DeerFlow](https:\u002F\u002Fgithub.com\u002Fbytedance\u002Fdeer-flow) | 字节开源 long-horizon SuperAgent harness，适合研究 Deep Research、报告、slides、网页、图像和视频生成。 |\n| [smolagents](https:\u002F\u002Fgithub.com\u002Fhuggingface\u002Fsmolagents) | Hugging Face 轻量 agent 框架，CodeAgent 思路值得研究。 |\n| [LangGraph](https:\u002F\u002Flangchain-ai.github.io\u002Flanggraph\u002F) | 状态图和可控 agent 编排，仍然值得作为工程基础学习。 |\n| [Qwen-Agent](https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen-Agent) | 国产模型生态里的工具调用、RAG、MCP agent 框架。 |\n| [Pydantic AI](https:\u002F\u002Fpydantic.dev\u002Fdocs\u002Fai\u002Fcore-concepts\u002Fagent\u002F) | 类型安全和结构化输出，适合生产 Python agent。 |\n\n### Legacy Or Optional Frameworks\n\n这些项目仍有参考价值，但不建议作为当前学习主线。\n\n| Framework | Note |\n| --- | --- |\n| [CrewAI](https:\u002F\u002Fdocs.crewai.com\u002F) | 可了解 role\u002Ftask\u002Fcrew 抽象，但很多场景已经被更强的 coding agent \u002F harness 形态覆盖。 |\n| [AutoGen](https:\u002F\u002Fmicrosoft.github.io\u002Fautogen\u002F) | 多 agent 对话框架经典项目，适合了解历史和论文，不必重押。 |\n| [LangChain Agents](https:\u002F\u002Fdocs.langchain.com\u002F) | 生态仍重要，但建议重点转向 LangGraph 和具体工程模式。 |\n\n### Papers\n\n| Paper | Topic |\n| --- | --- |\n| [ReAct](https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.03629) | Reasoning + acting 的基础范式。 |\n| [Toolformer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2302.04761) | 模型学习何时调用工具。 |\n| [Reflexion](https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.11366) | 语言反馈和自我改进。 |\n| [Generative Agents](https:\u002F\u002Farxiv.org\u002Fabs\u002F2304.03442) | 记忆、反思、规划驱动的模拟 agent。 |\n| [Voyager](https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.16291) | 开放世界中的长期学习 agent。 |\n| [AutoGen](https:\u002F\u002Farxiv.org\u002Fabs\u002F2308.08155) | 多 agent 对话框架。 |\n| [AgentBench](https:\u002F\u002Farxiv.org\u002Fabs\u002F2308.03688) | Agent 能力评测。 |\n| [WebArena](https:\u002F\u002Farxiv.org\u002Fabs\u002F2307.13854) | 真实网页环境下的 agent benchmark。 |\n| [SWE-bench](https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.06770) | 真实 GitHub issue 修复评测。 |\n| [SWE-agent](https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.15793) | 软件工程 agent 的 agent-computer interface。 |\n| [Dive into Claude Code](https:\u002F\u002Farxiv.org\u002Fabs\u002F2604.14228) | 从系统设计角度分析 Claude Code 这类 coding agent 的 harness、权限、压缩和扩展机制。 |\n| [AI Harness Engineering](https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.13357) | 把 harness 作为 agent 能力来源来研究。 |\n| [Configuring Agentic AI Coding Tools](https:\u002F\u002Farxiv.org\u002Fabs\u002F2602.14690) | 对 Claude Code、Codex、Gemini、Cursor 等 coding tools 的配置机制分析。 |\n| [Your Agent, Their Asset](https:\u002F\u002Farxiv.org\u002Fabs\u002F2604.04759) | OpenClaw 等本地 agent 的真实安全风险分析。 |\n\n### GitHub Repositories\n\n| Repo | Why It Is Useful |\n| --- | --- |\n| [datawhalechina\u002Fhello-agents](https:\u002F\u002Fgithub.com\u002Fdatawhalechina\u002Fhello-agents) | 中文智能体系统教程，从零开始构建智能体。 |\n| [shareAI-lab\u002Flearn-claude-code](https:\u002F\u002Fgithub.com\u002FshareAI-lab\u002Flearn-claude-code) | Bash is all you need，从 0 到 1 学 Claude Code-like agent harness。 |\n| [shareAI-lab\u002Fclaw0](https:\u002F\u002Fgithub.com\u002FshareAI-lab\u002Fclaw0) | 从 0 到 1 学 OpenClaw-like gateway，10 个渐进章节覆盖 agent loop 到 concurrency。 |\n| [openclaw\u002Fopenclaw](https:\u002F\u002Fgithub.com\u002Fopenclaw\u002Fopenclaw) | 研究本地个人 agent、skills、长运行任务、系统工具和权限边界。 |\n| [NousResearch\u002Fhermes-agent](https:\u002F\u002Fgithub.com\u002FNousResearch\u002Fhermes-agent) | 研究自托管 agent、长期记忆、skills、toolsets、多平台消息网关。 |\n| [ttguy0707\u002FCyberClaw](https:\u002F\u002Fgithub.com\u002Fttguy0707\u002FCyberClaw) | 研究透明 agent、安全审计、两段式执行、双水位记忆和心跳任务。 |\n| [bytedance\u002Fdeer-flow](https:\u002F\u002Fgithub.com\u002Fbytedance\u002Fdeer-flow) | Deep Research \u002F long-horizon agent，适合学习搜索、报告生成、沙箱、memory、skills、subagents、message gateway。 |\n| [NirDiamant\u002FGenAI_Agents](https:\u002F\u002Fgithub.com\u002FNirDiamant\u002FGenAI_Agents) | 综合教程库，适合横向比较 ReAct、Plan-and-Execute、Multi-Agent 等实现方式。 |\n| [NirDiamant\u002Fagents-towards-production](https:\u002F\u002Fgithub.com\u002FNirDiamant\u002Fagents-towards-production) | production-grade agent 教程库，适合补工程化组件。 |\n| [huggingface\u002Fsmolagents](https:\u002F\u002Fgithub.com\u002Fhuggingface\u002Fsmolagents) | 轻量 CodeAgent 风格框架，适合研究“让模型写代码执行动作”的范式。 |\n| [openai\u002Fcodex](https:\u002F\u002Fgithub.com\u002Fopenai\u002Fcodex) | OpenAI 开源 coding agent CLI，适合研究 sandbox、approval、CLI 产品形态。 |\n| [opencode-ai\u002Fopencode](https:\u002F\u002Fgithub.com\u002Fopencode-ai\u002Fopencode) | 终端优先的开源 coding agent，适合对照 Claude Code \u002F Codex。 |\n| [langchain-ai\u002Flanggraph](https:\u002F\u002Fgithub.com\u002Flangchain-ai\u002Flanggraph) | 状态图和可控 agent 编排。 |\n| [openai\u002Fopenai-agents-python](https:\u002F\u002Fgithub.com\u002Fopenai\u002Fopenai-agents-python) | OpenAI Agents SDK Python。 |\n| [QwenLM\u002FQwen-Agent](https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen-Agent) | Qwen 生态 agent 框架。 |\n| [browser-use\u002Fbrowser-use](https:\u002F\u002Fgithub.com\u002Fbrowser-use\u002Fbrowser-use) | 浏览器 agent 实战。 |\n| [bytedance\u002FUI-TARS-desktop](https:\u002F\u002Fgithub.com\u002Fbytedance\u002FUI-TARS-desktop) | 多模态桌面 agent stack，适合研究视觉 UI 操作。 |\n| [SWE-agent\u002FSWE-agent](https:\u002F\u002Fgithub.com\u002FSWE-agent\u002FSWE-agent) | 软件工程 agent。 |\n| [All-Hands-AI\u002FOpenHands](https:\u002F\u002Fgithub.com\u002FAll-Hands-AI\u002FOpenHands) | 开源 coding agent。 |\n| [microsoft\u002Fai-agents-for-beginners](https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002Fai-agents-for-beginners) | 系统化入门课程。 |\n| [jjyaoao\u002FHelloAgents](https:\u002F\u002Fgithub.com\u002Fjjyaoao\u002FHelloAgents) | 基于 OpenAI 原生 API 的生产级多智能体框架，覆盖 ToolResponse、上下文工程、会话持久化、子代理、TraceLogger 等。 |\n\n### Thoughtful Blogs\n\n| Blog | Why It Is Useful |\n| --- | --- |\n| [Lilian Weng: LLM Powered Autonomous Agents](https:\u002F\u002Flilianweng.github.io\u002Fposts\u002F2023-06-23-agent\u002F) | 经典长文，系统整理 agent 架构、记忆、规划和工具使用。 |\n| [Simon Willison: AI\u002FLLM writing](https:\u002F\u002Fsimonwillison.net\u002Ftags\u002Fllms\u002F) | 非常务实的 LLM 工程观察，适合补工程直觉。 |\n| [LangChain Blog](https:\u002F\u002Fblog.langchain.com\u002F) | LangGraph、LangSmith、agent 工程实践。 |\n| [Google Developers Blog: ADK](https:\u002F\u002Fdevelopers.googleblog.com\u002Fagent-development-kit-easy-to-build-multi-agent-applications\u002F) | Google ADK 官方发布文章。 |\n\n### Claude Code Study Path\n\nClaude Code 是当前最值得拆解的 coding agent 产品之一。推荐按“官方文档 -> 复刻项目 -> 架构解析 -> 工程对照”的顺序学习：\n\n- 读官方文档，理解 hooks、subagents、MCP、GitHub Actions、permissions。\n- 读公开分析论文，抽象出 agent harness 的设计空间。\n- 跟随 [learn-claude-code](https:\u002F\u002Fgithub.com\u002FshareAI-lab\u002Flearn-claude-code) 从零复刻核心机制。\n- 对照 [hello-agents](https:\u002F\u002Fgithub.com\u002Fdatawhalechina\u002Fhello-agents)、[OpenClaw](https:\u002F\u002Fgithub.com\u002Fopenclaw\u002Fopenclaw)、[Hermes Agent](https:\u002F\u002Fgithub.com\u002FNousResearch\u002Fhermes-agent) 这些开源项目，学习工程化实现。\n\n| Resource | Focus |\n| --- | --- |\n| [Claude Code Common Workflows](https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Ftutorials) | 官方工作流教程，适合学日常使用和项目协作。 |\n| [Claude Code Overview](https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Foverview) | 官方入口，理解产品能力和命令行工作方式。 |\n| [learn-claude-code](https:\u002F\u002Fgithub.com\u002FshareAI-lab\u002Flearn-claude-code) | 从零复刻 Claude Code-like agent，适合学 harness 最小实现。 |\n| [Claude Code 源码解析](https:\u002F\u002Fclaudecoding.dev\u002F) | 中文架构解读，适合按模块理解实现思路。 |\n| [Claude Code 源码分析地图](https:\u002F\u002Fcode.claudecn.com\u002F) | 用地图方式拆解 agent loop、工具、命令和多代理协作。 |\n| [Dive into Claude Code](https:\u002F\u002Farxiv.org\u002Fabs\u002F2604.14228) | 从研究视角总结 Claude Code 的设计空间和 agent harness 模式。 |\n| [Agentic Education](https:\u002F\u002Farxiv.org\u002Fabs\u002F2604.17460) | 用 Claude Code 学 Claude Code 的交互式课程思路。 |\n\n## Learning Principles\n\n- Build first, then read deeper.\n- Prefer small reliable agents over impressive demos.\n- Use tools with strict schemas.\n- Add evals before you add more agents.\n- Trace every important run.\n- Treat multi-agent as a coordination problem.\n- Keep humans in the loop for risky actions.\n- Respect platform rules, copyrights, and data access boundaries.\n\n## Contributing\n\nGood contributions are welcome. Please prefer:\n\n- official docs and official engineering blogs;\n- high-quality papers and benchmarks;\n- open-source repositories with runnable code;\n- serious technical blogs with original insight;\n- small projects that help learners practice one specific skill.\n\nPlease avoid:\n\n- copied platform posts;\n- course ads without substantial content;\n- private or paywalled material;\n- scraping-oriented content that bypasses platform rules.\n","Agent Learning Hub 是一个为希望构建有用且可靠的AI代理的学习路线和资源库。该项目通过整理社区优秀分享、官方博客、论文及开源项目，形成了一份详尽的AI Agent学习待办清单，特别强调了从基础理解到实际应用的系统化学习路径。技术特点包括但不限于：构建最小代理循环、工具使用与调用、检索增强生成（RAG）、记忆管理等。适用于想要深入掌握AI代理开发技术的初学者以及希望提升特定技能如编码代理或个人代理工程的专业人士。无论你是刚开始接触AI代理领域还是已经在寻找具体项目灵感，都能在这里找到适合自己的学习材料和发展方向。",2,"2026-06-11 03:55:32","CREATED_QUERY"]