[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"pack-detail-multi-agent-frameworks-zh":3,"seo:pack:multi-agent-frameworks:zh":77},{"code":4,"message":5,"data":6},200,"操作成功",{"pack":7},{"slug":8,"icon":9,"tone":10,"status":11,"status_label":12,"title":13,"description":14,"items":15,"install_cmd":76},"multi-agent-frameworks","🤝","#F43F5E","stable","稳定","多 Agent 框架","CAMEL \u002F LangGraph \u002F DeepAgents \u002F GPT Researcher — 把多个 agent 编成一队上生产的框架。",[16,28,35,42,50,60,68],{"id":17,"uuid":18,"slug":19,"title":20,"description":21,"author_name":22,"view_count":23,"vote_count":24,"lang_type":25,"type":26,"type_label":27},295,"23732313-ea97-4319-b7a5-19dcddd7e97c","camel-multi-agent-framework-scale-23732313","CAMEL — Multi-Agent Framework at Scale","CAMEL is a multi-agent framework for studying scaling laws of AI agents. 16.6K+ GitHub stars. Up to 1M agents, RAG, memory systems, data generation. Apache 2.0.","Script Depot",628,0,"en","skill","Skill",{"id":4,"uuid":29,"slug":30,"title":31,"description":32,"author_name":33,"view_count":34,"vote_count":24,"lang_type":25,"type":26,"type_label":27},"cc1a6ed2-0d82-4379-94f4-15632b4d4967","langgraph-build-stateful-ai-agents-graphs-cc1a6ed2","LangGraph — Build Stateful AI Agents as Graphs","LangChain framework for building resilient, stateful AI agents as graphs. Supports cycles, branching, persistence, human-in-the-loop, and streaming. 28K+ stars.","LangChain",566,{"id":36,"uuid":37,"slug":38,"title":39,"description":40,"author_name":33,"view_count":41,"vote_count":24,"lang_type":25,"type":26,"type_label":27},601,"ac820f80-41ff-4eaa-b3b1-da27653bd7a5","deepagents-multi-step-agent-framework-langchain-ac820f80","DeepAgents — Multi-Step Agent Framework by LangChain","Agent harness built on LangGraph by the LangChain team. Features planning tools, filesystem backend, and sub-agent spawning for complex multi-step tasks like codebase refactoring. 16,500+ stars.",431,{"id":43,"uuid":44,"slug":45,"title":46,"description":47,"author_name":48,"view_count":49,"vote_count":24,"lang_type":25,"type":26,"type_label":27},25,"23330210-b26a-4d97-ad97-1735c203eaa6","gpt-researcher-autonomous-research-report-agent-23330210","GPT Researcher — Autonomous Research Report Agent","AI agent that generates detailed research reports from a single query. Searches multiple sources, synthesizes findings, and cites references.","TokRepo精选",1749,{"id":51,"uuid":52,"slug":53,"title":54,"description":55,"author_name":56,"view_count":57,"vote_count":24,"lang_type":25,"type":58,"type_label":59},719,"dedbb70b-7ebd-4987-9ba3-a883f45e8b5a","goose-ai-developer-agent-block-dedbb70b","Goose — AI Developer Agent by Block","Open-source AI developer agent by Block (Square). Goose automates coding tasks with extensible toolkits, session memory, and MCP server support in your terminal.","Block",406,"script","Script",{"id":61,"uuid":62,"slug":63,"title":64,"description":65,"author_name":66,"view_count":67,"vote_count":24,"lang_type":25,"type":26,"type_label":27},2221,"34ff4f3b-1660-4953-9053-7b3fac036c17","claude-flow-multi-agent-orchestration-claude-code-34ff4f3b","Claude-Flow — Multi-Agent Orchestration for Claude Code","Layers swarm and hive-mind multi-agent orchestration on top of Claude Code with 64 specialized agents, SQLite memory, and parallel execution.","Skill Factory",404,{"id":69,"uuid":70,"slug":71,"title":72,"description":73,"author_name":74,"view_count":75,"vote_count":24,"lang_type":25,"type":58,"type_label":59},683,"38035d0b-f942-4bf2-bbad-be9d4f719c00","openai-agents-sdk-build-multi-agent-systems-python-38035d0b","OpenAI Agents SDK — Build Multi-Agent Systems in Python","Official OpenAI Python SDK for building multi-agent systems with handoffs, guardrails, and tracing. Agents delegate to specialists, enforce safety rules, and produce observable traces. 8,000+ stars.","OpenAI",263,"tokrepo install pack\u002Fmulti-agent-frameworks",{"pageType":78,"pageKey":8,"locale":79,"title":80,"metaDescription":81,"h1":13,"tldr":82,"bodyMarkdown":83,"faq":84,"schema":100,"internalLinks":110,"citations":123,"wordCount":136,"generatedAt":137},"pack","zh","多 Agent 框架：CAMEL \u002F LangGraph \u002F DeepAgents 七件套","CAMEL \u002F LangGraph \u002F DeepAgents \u002F GPT Researcher —— 2026 团队真在生产用的七个多 agent 编排框架。TokRepo 一条命令装齐。","七个多 agent 编排框架 —— CAMEL \u002F LangGraph \u002F DeepAgents \u002F GPT Researcher + 三个角色扮演与研究模板 —— 都能扛住生产。TokRepo 一条命令装好。","## 这个 pack 装了什么\n\n这个 pack 收齐了**七个多 agent 框架**，是 2026 年团队真在生产里跑的那一批，不是 Twitter 上看着炫但一上量就崩的 demo。四个是头部框架，三个是基于它们封装的研究 \u002F 角色模板。\n\n| # | 资产 | 类型 | 适合场景 |\n|---|---|---|---|\n| 1 | LangGraph | 状态化框架 | 带 checkpoint 的生产图编排 |\n| 2 | CAMEL | 角色扮演框架 | Agent 间对话，学术级 |\n| 3 | DeepAgents | 研究框架 | 长任务规划 + sub-agent 派生 |\n| 4 | GPT Researcher | 应用 agent | 给主题、出研究报告 |\n| 5 | 研究员 swarm | 模板 | 并行研究的 CAMEL 角色 |\n| 6 | 评审-执行对 | 模板 | 一个 agent 做事，一个 agent 评审，纠错 |\n| 7 | 分级 planner | 模板 | 经理派给工人的模式带预算 |\n\n## 这个 pack 为什么重要\n\n单 agent 是个聊天循环。多 agent 是个系统 —— 跟所有系统一样，要扛住真实负载就得有结构（状态机、队列、重试）。这里四个框架挑的是真能用的结构。三个模板告诉你最常见用例怎么接起来。\n\n四个框架各下了不同的抽象赌注：\n\n- **LangGraph** 把编排当作状态图。你声明节点（agent \u002F 工具）和边（什么时候跳转），LangGraph 负责 checkpointing 让 30 分钟的任务崩了能续。最接近生产事实标准的那个。\n- **CAMEL** 聚焦显式角色的 agent 间对话。两个 agent 演「用户」「助手」或「研究负责人」「写手」，对话直到目标达成。可复现性和学术 benchmark 强。\n- **DeepAgents** 为长任务而生。顶层 agent 做规划，把子任务派给派生的 sub-agent，每个 sub-agent 有自己的 context window。专门避开「一个巨大 context」的失败模式。\n- **GPT Researcher** 是应用案例。你给它一个研究问题，它跑一个 sub-agent swarm 收证据，输出带引用的长文报告。既是工具也是参考架构。\n\n## 一条命令装齐\n\n```bash\n# 装整个 pack\ntokrepo install pack\u002Fmulti-agent-frameworks\n\n# 或者一个一个装\ntokrepo install langgraph\ntokrepo install camel\ntokrepo install deepagents\ntokrepo install gpt-researcher\n```\n\nTokRepo CLI 把每个框架的适配器装进你的 AI 工具 —— Claude Code subagent 进 `.claude\u002Fagents\u002F`，Cursor 规则进 `.cursor\u002Frules\u002F`，Codex CLI 进 AGENTS.md。底层库自己 pip \u002F npm 装；TokRepo 接好 prompt 让你的 CLI 知道什么时候触发。\n\n## 常见坑\n\n- **别忘了预算**。多 agent 任务能指数扇出 —— 一个 planner 派 5 个工人每个再派 5 个子任务，token 烧 25 倍。永远封顶深度和最大派生数。DeepAgents 内建这个；LangGraph 和 CAMEL 你自己设。\n- **别天真地跨线程共用 LLM 客户端**。多数 SDK 高并发下不完全线程安全。用进程级池或带界限的 async（如 asyncio.Semaphore(8)）。\n- **追踪一切**。没 trace 的多 agent 调试根本做不动。这个 pack 配 LLM 可观测性 pack 一起 —— Langfuse 和 AgentOps 都有 LangGraph 一等公民集成。\n- **小心角色漂移**。CAMEL 风格对话里，agent 第 8-10 轮经常忘了自己是谁。每 N 轮加 system 提醒，或在每条消息钉住角色。\n- **多 agent ≠ 更好**。先试单 Claude Sonnet 4.5 + 扩展思考，再考虑多 agent 系统。2025 年 Anthropic 多 agent 研究博客发现，60% 人们丢给多 agent 的任务，单 agent + 工具就能搞定。\n\n## 这个 pack 不够用的时候\n\n多 agent 在能并行的子问题上闪光（研究、代码评审、跨主题内容生成）。但在以下场景失利：\n\n- **顺序、深度状态化的任务**。整库重构是单 agent 的活 —— 切分到多个 agent 反而协调开销大于收益。\n- **延迟敏感的工作流**。每次 agent 间跳转都加一个回合。SLA 5 秒以内的，留单 agent。\n- **成本敏感的工作流**。多 agent 同任务通常 3-10 倍单 agent 成本。难题求质量值得；「总结这封邮件」不值。\n\n正确的上手路径：先用 GPT Researcher 当最简成品例子，等需要自己写编排时再升级 LangGraph 或 DeepAgents。",[85,88,91,94,97],{"q":86,"a":87},"LangGraph 免费吗？","是。LangGraph 开源 MIT 许可，你只付 LLM token 钱。有付费 LangGraph Cloud 提供托管部署带 checkpoint 和追踪，但开源库功能完整，无付费层也能上生产。CAMEL \u002F DeepAgents \u002F GPT Researcher 也都是 OSS。",{"q":89,"a":90},"Cursor \u002F Codex CLI 能用吗？","框架都是语言级 Python 库，不是 Claude Code 专属。任何能跑 Python 工具的 agent CLI 都能驱动。TokRepo CLI 给每个工具装对应接线 —— Codex CLI 出 AGENTS.md 指令说什么时候调用，Cursor 加规则。底层 Python 安装不变。",{"q":92,"a":93},"LangGraph 跟 CAMEL 怎么选？","LangGraph 结构优先：你画一个状态机，agent 进去填槽。CAMEL 对话优先：你分配角色让 agent 自由对话。LangGraph 在生产可靠性和 checkpoint 上赢；CAMEL 在研究、仿真、对话本身就是产物的场景赢。很多生产设置用 LangGraph 编排，CAMEL 处理具体对话任务。",{"q":95,"a":96},"跟记忆层 pack 有啥区别？","记忆是「agent 跨会话记住什么」。多 agent 是「单个任务里多 agent 怎么协调」。两者正交：多 agent 系统经常需要共享记忆层（Mem0 \u002F Zep）让工人不必重发现 planner 早就知道的事实。要做正经东西建议两个 pack 都装。",{"q":98,"a":99},"什么时候*不*该上多 agent 框架？","任务顺序又有状态时（重构这个文件）、延迟敏感（聊天 UI 3 秒内）、或简单到一次 Claude \u002F GPT 调用就够。Anthropic 自己的多 agent 研究博客指出，单 agent + 扩展思考在成本上打败多数多 agent 配置。任务天然并行（研究多源）或需要不同专家角色时再上多 agent。",{"@context":101,"@type":102,"name":103,"description":104,"numberOfItems":105,"publisher":106},"https:\u002F\u002Fschema.org","CollectionPage","Multi-Agent Frameworks","CAMEL, LangGraph, DeepAgents, GPT Researcher — frameworks for orchestrating teams of agents in production.",7,{"@type":107,"name":108,"url":109},"Organization","TokRepo","https:\u002F\u002Ftokrepo.com",[111,115,119],{"url":112,"anchor":113,"reason":114},"\u002Fzh\u002Fpacks\u002Fagent-memory-layer","Agent 记忆层","swarm 中跨 agent 共享记忆",{"url":116,"anchor":117,"reason":118},"\u002Fzh\u002Fpacks\u002Fllm-observability","LLM 可观测性","追踪多 agent 调用找瓶颈",{"url":120,"anchor":121,"reason":122},"\u002Fzh\u002Ftools\u002Fclaude-code","Claude Code","编排这些框架的常见宿主",[124,128,132],{"claim":125,"source_name":126,"source_url":127},"LangGraph is the official stateful orchestration library from the LangChain team","langchain-ai\u002Flanggraph on GitHub","https:\u002F\u002Fgithub.com\u002Flangchain-ai\u002Flanggraph",{"claim":129,"source_name":130,"source_url":131},"CAMEL is one of the earliest multi-agent role-playing frameworks, with active research output","camel-ai\u002Fcamel on GitHub","https:\u002F\u002Fgithub.com\u002Fcamel-ai\u002Fcamel",{"claim":133,"source_name":134,"source_url":135},"GPT Researcher autonomously runs research tasks and produces long-form reports","assafelovic\u002Fgpt-researcher on GitHub","https:\u002F\u002Fgithub.com\u002Fassafelovic\u002Fgpt-researcher",547,"2026-05-02T15:00:00Z"]