[{"data":1,"prerenderedAt":857},["ShallowReactive",2],{"blog-post-zh-cn-claude-code-agent-teams-guide":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"tags":11,"body":14,"_type":851,"_id":852,"_source":853,"_file":854,"_stem":855,"_extension":856},"\u002Fzh-cn\u002Fblog\u002Fclaude-code-agent-teams-guide","blog",false,"","Claude Code Agent Teams 实用入门：让一群 AI 在终端里组队干活","详解 Claude Code 的 Agent Teams：团队负责人、点对点信箱、共享任务列表如何协作，适合什么场景，如何开启，以及必须了解的注意事项。","2026-07-18",[12,13],"claude-code","agent-frameworks",{"type":15,"children":16,"toc":837},"root",[17,44,49,56,75,80,105,119,139,145,150,243,250,264,269,281,286,291,304,310,324,347,366,371,376,426,438,444,465,477,482,495,570,575,607,612,696,701,706,711,772,778,814,818,831],{"type":18,"tag":19,"props":20,"children":21},"element","p",{},[22,25,34,36,42],{"type":23,"value":24},"text","2026 年 2 月初，Anthropic 发布 ",{"type":18,"tag":26,"props":27,"children":31},"a",{"href":28,"rel":29},"https:\u002F\u002Fwww.anthropic.com\u002Fnews\u002Fclaude-opus-4-6",[30],"nofollow",[32],{"type":23,"value":33},"Claude Opus 4.6",{"type":23,"value":35},"，新闻焦点都在模型本身。但对关注\"AI 团队\"这件事的人来说，真正值得留意的功能藏在 Claude Code 里——",{"type":18,"tag":37,"props":38,"children":39},"strong",{},[40],{"type":23,"value":41},"Agent Teams（智能体团队）",{"type":23,"value":43},"。这是一项研究预览（research preview）功能，允许一个 Claude Code 会话生成并协调多个其他 Claude Code 会话并行工作。你可以在终端里亲眼看着一支小型 AI 工程团队组建起来：分工、认领任务、互相争论、汇总结论。",{"type":18,"tag":19,"props":45,"children":46},{},[47],{"type":23,"value":48},"这篇文章是一份实用入门：Agent Teams 到底是什么，各个部件怎么配合，哪些场景真正有用、哪些不适合，以及怎么把它打开。",{"type":18,"tag":50,"props":51,"children":53},"h2",{"id":52},"agent-teams-是什么以及不是什么",[54],{"type":23,"value":55},"Agent Teams 是什么（以及不是什么）",{"type":18,"tag":19,"props":57,"children":58},{},[59,61,66,68,73],{"type":23,"value":60},"Agent Teams 是直接内置在 Claude Code 里的实验性多智能体编排能力。你正在输入的那个会话会成为",{"type":18,"tag":37,"props":62,"children":63},{},[64],{"type":23,"value":65},"团队负责人（team lead）",{"type":23,"value":67},"；按你的指示，它会生成若干",{"type":18,"tag":37,"props":69,"children":70},{},[71],{"type":23,"value":72},"队友（teammate）",{"type":23,"value":74},"——每个队友都是一个完整、独立的 Claude Code 实例，拥有自己的上下文窗口、自己的工具权限、自己对项目的独立视角。",{"type":18,"tag":19,"props":76,"children":77},{},[78],{"type":23,"value":79},"和以往\"派个帮手\"的模式相比，它有两个关键区别：",{"type":18,"tag":81,"props":82,"children":83},"ol",{},[84,95],{"type":18,"tag":85,"props":86,"children":87},"li",{},[88,93],{"type":18,"tag":37,"props":89,"children":90},{},[91],{"type":23,"value":92},"队友之间可以直接对话",{"type":23,"value":94},"，不必事事经过负责人，通信是点对点的。",{"type":18,"tag":85,"props":96,"children":97},{},[98,103],{"type":18,"tag":37,"props":99,"children":100},{},[101],{"type":23,"value":102},"工作通过共享任务列表来协调",{"type":23,"value":104},"，队友可以自己认领任务，而不是负责人事无巨细地派活。",{"type":18,"tag":19,"props":106,"children":107},{},[108,110,117],{"type":23,"value":109},"这个功能明确处于实验阶段：默认关闭，需要环境变量开启，",{"type":18,"tag":26,"props":111,"children":114},{"href":112,"rel":113},"https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Fagent-teams",[30],[115],{"type":23,"value":116},"官方文档",{"type":23,"value":118},"开篇就列出了已知限制。把它当作一种工作方式的预览，而不是成品。",{"type":18,"tag":19,"props":120,"children":121},{},[122,124,129,131,137],{"type":23,"value":123},"同样需要说清楚它",{"type":18,"tag":37,"props":125,"children":126},{},[127],{"type":23,"value":128},"不是",{"type":23,"value":130},"什么：Agent Teams 是一个交互式的开发工作流工具。它运行在你的终端会话里，期望你在旁边观察和把舵，团队随会话而生、随会话而灭。它不是生产级智能体框架——没有部署方案，没有持久化编排，也没有把团队跑成服务的 API。如果你要找的是后者，请看我们的",{"type":18,"tag":26,"props":132,"children":134},{"href":133},"\u002Fzh-CN\u002Fblog\u002Fbest-multi-agent-frameworks-2026",[135],{"type":23,"value":136},"多智能体框架横评",{"type":23,"value":138},"。",{"type":18,"tag":50,"props":140,"children":142},{"id":141},"工作原理四个核心部件",[143],{"type":23,"value":144},"工作原理：四个核心部件",{"type":18,"tag":19,"props":146,"children":147},{},[148],{"type":23,"value":149},"按照文档的描述，一个智能体团队由四部分组成：",{"type":18,"tag":151,"props":152,"children":153},"table",{},[154,174],{"type":18,"tag":155,"props":156,"children":157},"thead",{},[158],{"type":18,"tag":159,"props":160,"children":161},"tr",{},[162,169],{"type":18,"tag":163,"props":164,"children":166},"th",{"align":165},"left",[167],{"type":23,"value":168},"部件",{"type":18,"tag":163,"props":170,"children":171},{"align":165},[172],{"type":23,"value":173},"作用",{"type":18,"tag":175,"props":176,"children":177},"tbody",{},[178,195,211,227],{"type":18,"tag":159,"props":179,"children":180},{},[181,190],{"type":18,"tag":182,"props":183,"children":184},"td",{"align":165},[185],{"type":18,"tag":37,"props":186,"children":187},{},[188],{"type":23,"value":189},"团队负责人",{"type":18,"tag":182,"props":191,"children":192},{"align":165},[193],{"type":23,"value":194},"你的主会话。生成队友、创建任务、汇总结果",{"type":18,"tag":159,"props":196,"children":197},{},[198,206],{"type":18,"tag":182,"props":199,"children":200},{"align":165},[201],{"type":18,"tag":37,"props":202,"children":203},{},[204],{"type":23,"value":205},"队友",{"type":18,"tag":182,"props":207,"children":208},{"align":165},[209],{"type":23,"value":210},"独立的 Claude Code 实例，各有自己的上下文窗口",{"type":18,"tag":159,"props":212,"children":213},{},[214,222],{"type":18,"tag":182,"props":215,"children":216},{"align":165},[217],{"type":18,"tag":37,"props":218,"children":219},{},[220],{"type":23,"value":221},"任务列表",{"type":18,"tag":182,"props":223,"children":224},{"align":165},[225],{"type":23,"value":226},"共享的工作项，队友认领并完成",{"type":18,"tag":159,"props":228,"children":229},{},[230,238],{"type":18,"tag":182,"props":231,"children":232},{"align":165},[233],{"type":18,"tag":37,"props":234,"children":235},{},[236],{"type":23,"value":237},"信箱",{"type":18,"tag":182,"props":239,"children":240},{"align":165},[241],{"type":23,"value":242},"任意两个智能体之间的点对点消息通道",{"type":18,"tag":244,"props":245,"children":247},"h3",{"id":246},"信箱mailbox",[248],{"type":23,"value":249},"信箱（Mailbox）",{"type":18,"tag":19,"props":251,"children":252},{},[253,255,262],{"type":23,"value":254},"每个智能体都有一个信箱——实现上就是 ",{"type":18,"tag":256,"props":257,"children":259},"code",{"className":258},[],[260],{"type":23,"value":261},"~\u002F.claude\u002Fteams\u002F{team-name}\u002Finboxes\u002F",{"type":23,"value":263}," 目录下的一个 JSON 文件——任何队友都可以按名字给另一个队友发消息。听起来只是底层管道，但这恰恰是架构上最有意思的地方：发现、异议、请求都在智能体之间直接流转，不需要负责人当中转站。负责人也不用轮询进度——消息自动送达，队友干完活闲下来时会主动通知负责人。",{"type":18,"tag":244,"props":265,"children":267},{"id":266},"共享任务列表",[268],{"type":23,"value":266},{"type":18,"tag":19,"props":270,"children":271},{},[272,274,279],{"type":23,"value":273},"负责人把工作拆成任务，任务有三种状态（待办、进行中、已完成），还可以声明依赖关系——依赖未完成的任务无法被认领，直到阻塞解除。任务可以由负责人显式指派，也可以由队友",{"type":18,"tag":37,"props":275,"children":276},{},[277],{"type":23,"value":278},"自主认领",{"type":23,"value":280},"：做完一个，自己去拿下一个无人认领、无阻塞的任务。认领机制用文件锁防止两个队友抢到同一个任务。",{"type":18,"tag":19,"props":282,"children":283},{},[284],{"type":23,"value":285},"正是这套机制让系统更像一个\"团队\"，而不是\"调度器带一群工人\"。你可以给负责人十五个独立任务和三个队友，工作会自己分配开。",{"type":18,"tag":244,"props":287,"children":289},{"id":288},"队友知道什么",[290],{"type":23,"value":288},{"type":18,"tag":19,"props":292,"children":293},{},[294,296,302],{"type":23,"value":295},"新生成的队友会加载与普通会话相同的项目上下文——",{"type":18,"tag":256,"props":297,"children":299},{"className":298},[],[300],{"type":23,"value":301},"CLAUDE.md",{"type":23,"value":303},"、MCP 服务器、技能（skills）——外加负责人为它写的生成提示词（spawn prompt）。但它拿不到负责人的对话历史。实际影响是：任务相关的背景信息必须写进生成提示词里，因为你向负责人解释问题时，队友\"不在场\"。",{"type":18,"tag":50,"props":305,"children":307},{"id":306},"agent-teams-与子智能体subagents的区别",[308],{"type":23,"value":309},"Agent Teams 与子智能体（Subagents）的区别",{"type":18,"tag":19,"props":311,"children":312},{},[313,315,322],{"type":23,"value":314},"Claude Code 早就有",{"type":18,"tag":26,"props":316,"children":319},{"href":317,"rel":318},"https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Fsub-agents",[30],[320],{"type":23,"value":321},"子智能体",{"type":23,"value":323},"功能，两者的区别很重要：",{"type":18,"tag":325,"props":326,"children":327},"ul",{},[328,337],{"type":18,"tag":85,"props":329,"children":330},{},[331,335],{"type":18,"tag":37,"props":332,"children":333},{},[334],{"type":23,"value":321},{"type":23,"value":336},"在你的会话内运行，做一件聚焦的事，把结果汇报给调用方，彼此之间从不通信。更便宜、更简单，适合\"去查一下，告诉我结论\"这类需求。",{"type":18,"tag":85,"props":338,"children":339},{},[340,345],{"type":18,"tag":37,"props":341,"children":342},{},[343],{"type":23,"value":344},"Agent Teams",{"type":23,"value":346}," 是完全独立的会话，共享任务列表、互发消息。更贵、能力也更强，适合需要讨论和并行分工的工作。",{"type":18,"tag":19,"props":348,"children":349},{},[350,352,357,359,364],{"type":23,"value":351},"文档里有条很实用的经验法则：只关心",{"type":18,"tag":37,"props":353,"children":354},{},[355],{"type":23,"value":356},"结果",{"type":23,"value":358},"时用子智能体；工人之间需要",{"type":18,"tag":37,"props":360,"children":361},{},[362],{"type":23,"value":363},"协作",{"type":23,"value":365},"时用团队。",{"type":18,"tag":50,"props":367,"children":369},{"id":368},"适合什么场景",[370],{"type":23,"value":368},{"type":18,"tag":19,"props":372,"children":373},{},[374],{"type":23,"value":375},"这个模式在\"并行探索真正有价值\"的地方最出彩：",{"type":18,"tag":325,"props":377,"children":378},{},[379,389,406,416],{"type":18,"tag":85,"props":380,"children":381},{},[382,387],{"type":18,"tag":37,"props":383,"children":384},{},[385],{"type":23,"value":386},"并行代码评审。",{"type":23,"value":388}," 对同一个 PR 生成三个视角不同的评审员——安全、性能、测试覆盖。每人只戴一副\"滤镜\"看得透彻，胜过一个智能体什么都想看却什么都只看一眼。",{"type":18,"tag":85,"props":390,"children":391},{},[392,397,399,404],{"type":18,"tag":37,"props":393,"children":394},{},[395],{"type":23,"value":396},"多假设并行排障。",{"type":23,"value":398}," 文档里最有意思的例子：生成几个调查员，各自持有不同的故障假设，并要求他们像科学辩论一样",{"type":18,"tag":37,"props":400,"children":401},{},[402],{"type":23,"value":403},"互相证伪",{"type":23,"value":405},"。串行排查容易锚定在第一个\"看起来合理\"的解释上；对抗式的并行调查更可能找到真正的根因。",{"type":18,"tag":85,"props":407,"children":408},{},[409,414],{"type":18,"tag":37,"props":410,"children":411},{},[412],{"type":23,"value":413},"新模块、新功能。",{"type":23,"value":415}," 边界清晰、可以按文件划分的工作——一人写 API、一人写前端、一人写测试，互不碰对方的文件。",{"type":18,"tag":85,"props":417,"children":418},{},[419,424],{"type":18,"tag":37,"props":420,"children":421},{},[422],{"type":23,"value":423},"调研与探索。",{"type":23,"value":425}," 让队友从不同角度评估一个库、一个架构选型或一条迁移路径，最后汇总。",{"type":18,"tag":19,"props":427,"children":428},{},[429,431,436],{"type":23,"value":430},"反过来，它",{"type":18,"tag":37,"props":432,"children":433},{},[434],{"type":23,"value":435},"不适合",{"type":23,"value":437},"：串行任务、依赖关系繁重的工作，以及任何多个智能体会改同一批文件的场景（两个队友写一个文件，结局就是互相覆盖）。这些情况下，单会话或子智能体既便宜又好用。",{"type":18,"tag":50,"props":439,"children":441},{"id":440},"一个真实数据点claude-code-review",[442],{"type":23,"value":443},"一个真实数据点：Claude Code Review",{"type":18,"tag":19,"props":445,"children":446},{},[447,449,454,456,463],{"type":23,"value":448},"这个模式能干真活的最有力证据出现在 2026 年 3 月：Anthropic 发布了 ",{"type":18,"tag":37,"props":450,"children":451},{},[452],{"type":23,"value":453},"Claude Code Review",{"type":23,"value":455},"，一个",{"type":18,"tag":26,"props":457,"children":460},{"href":458,"rel":459},"https:\u002F\u002Fwww.vktr.com\u002Fai-news\u002Fanthropic-launches-multi-agent-code-review-for-claude-code\u002F",[30],[461],{"type":23,"value":462},"基于多智能体架构",{"type":23,"value":464},"的自动化 PR 评审工具——多个专职评审智能体并行检查每个 Pull Request，再由一个验证环节过滤误报，最后把结论以行内评论的形式发出。",{"type":18,"tag":19,"props":466,"children":467},{},[468,470,475],{"type":23,"value":469},"Anthropic 公布的内部使用数据是：部署之后，收到实质性评审意见的 PR 占比",{"type":18,"tag":37,"props":471,"children":472},{},[473],{"type":23,"value":474},"从 16% 升到了 54%",{"type":23,"value":476},"。照例要打个折扣——这是 Anthropic 在自家代码库上测自家产品——但它确实是一个具体的、生产形态的结果，用的正是 Agent Teams 交互式提供给你的那套\"多个专职智能体并行、再汇总\"的模式。",{"type":18,"tag":50,"props":478,"children":480},{"id":479},"如何上手",[481],{"type":23,"value":479},{"type":18,"tag":19,"props":483,"children":484},{},[485,487,493],{"type":23,"value":486},"Agent Teams 默认关闭，通过环境变量开启，可以设在 shell 里，也可以写进 ",{"type":18,"tag":256,"props":488,"children":490},{"className":489},[],[491],{"type":23,"value":492},"settings.json",{"type":23,"value":494},"：",{"type":18,"tag":496,"props":497,"children":501},"pre",{"className":498,"code":499,"language":500,"meta":7,"style":7},"language-json shiki shiki-themes github-dark","{\n  \"env\": {\n    \"CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS\": \"1\"\n  }\n}\n","json",[502],{"type":18,"tag":256,"props":503,"children":504},{"__ignoreMap":7},[505,517,532,552,561],{"type":18,"tag":506,"props":507,"children":510},"span",{"class":508,"line":509},"line",1,[511],{"type":18,"tag":506,"props":512,"children":514},{"style":513},"--shiki-default:#E1E4E8",[515],{"type":23,"value":516},"{\n",{"type":18,"tag":506,"props":518,"children":520},{"class":508,"line":519},2,[521,527],{"type":18,"tag":506,"props":522,"children":524},{"style":523},"--shiki-default:#79B8FF",[525],{"type":23,"value":526},"  \"env\"",{"type":18,"tag":506,"props":528,"children":529},{"style":513},[530],{"type":23,"value":531},": {\n",{"type":18,"tag":506,"props":533,"children":535},{"class":508,"line":534},3,[536,541,546],{"type":18,"tag":506,"props":537,"children":538},{"style":523},[539],{"type":23,"value":540},"    \"CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS\"",{"type":18,"tag":506,"props":542,"children":543},{"style":513},[544],{"type":23,"value":545},": ",{"type":18,"tag":506,"props":547,"children":549},{"style":548},"--shiki-default:#9ECBFF",[550],{"type":23,"value":551},"\"1\"\n",{"type":18,"tag":506,"props":553,"children":555},{"class":508,"line":554},4,[556],{"type":18,"tag":506,"props":557,"children":558},{"style":513},[559],{"type":23,"value":560},"  }\n",{"type":18,"tag":506,"props":562,"children":564},{"class":508,"line":563},5,[565],{"type":18,"tag":506,"props":566,"children":567},{"style":513},[568],{"type":23,"value":569},"}\n",{"type":18,"tag":19,"props":571,"children":572},{},[573],{"type":23,"value":574},"之后用自然语言描述你想要的团队即可。官方文档的示例提示词是个不错的模板：",{"type":18,"tag":496,"props":576,"children":579},{"className":577,"code":578,"language":23,"meta":7,"style":7},"language-text shiki shiki-themes github-dark","我在设计一个帮开发者追踪代码库中 TODO 注释的 CLI 工具。\n生成三个队友从不同角度探索：一个负责用户体验，\n一个负责技术架构，一个专门唱反调。\n",[580],{"type":18,"tag":256,"props":581,"children":582},{"__ignoreMap":7},[583,591,599],{"type":18,"tag":506,"props":584,"children":585},{"class":508,"line":509},[586],{"type":18,"tag":506,"props":587,"children":588},{},[589],{"type":23,"value":590},"我在设计一个帮开发者追踪代码库中 TODO 注释的 CLI 工具。\n",{"type":18,"tag":506,"props":592,"children":593},{"class":508,"line":519},[594],{"type":18,"tag":506,"props":595,"children":596},{},[597],{"type":23,"value":598},"生成三个队友从不同角度探索：一个负责用户体验，\n",{"type":18,"tag":506,"props":600,"children":601},{"class":508,"line":534},[602],{"type":18,"tag":506,"props":603,"children":604},{},[605],{"type":23,"value":606},"一个负责技术架构，一个专门唱反调。\n",{"type":18,"tag":19,"props":608,"children":609},{},[610],{"type":23,"value":611},"负责人会填充任务列表、生成队友，并在他们完成后做汇总。有几件事值得尽早知道：",{"type":18,"tag":325,"props":613,"children":614},{},[615,641,651,661],{"type":18,"tag":85,"props":616,"children":617},{},[618,623,625,631,633,639],{"type":18,"tag":37,"props":619,"children":620},{},[621],{"type":23,"value":622},"显示模式。",{"type":23,"value":624}," 默认一切在主终端内运行（in-process）：队友出现在输入框下方的面板里，方向键选中、回车打开对方的会话记录并直接对话。想让每个队友独占一个窗格，可以用 tmux 或 iTerm2 的分屏模式（在 ",{"type":18,"tag":256,"props":626,"children":628},{"className":627},[],[629],{"type":23,"value":630},"~\u002F.claude\u002Fsettings.json",{"type":23,"value":632}," 里设 ",{"type":18,"tag":256,"props":634,"children":636},{"className":635},[],[637],{"type":23,"value":638},"\"teammateMode\": \"auto\"",{"type":23,"value":640},"）。",{"type":18,"tag":85,"props":642,"children":643},{},[644,649],{"type":18,"tag":37,"props":645,"children":646},{},[647],{"type":23,"value":648},"可以直接和任何队友说话。",{"type":23,"value":650}," 每个队友都是完整会话——纠正它的思路、追问细节、叫停它，都不必经过负责人。",{"type":18,"tag":85,"props":652,"children":653},{},[654,659],{"type":18,"tag":37,"props":655,"children":656},{},[657],{"type":23,"value":658},"计划审批。",{"type":23,"value":660}," 对高风险工作，可以要求队友先出计划：队友在只读模式下工作，直到负责人批准其方案；你还可以给负责人定审批标准（比如\"只批含测试覆盖的计划\"）。",{"type":18,"tag":85,"props":662,"children":663},{},[664,669,671,678,680,686,688,694],{"type":18,"tag":37,"props":665,"children":666},{},[667],{"type":23,"value":668},"质量门禁。",{"type":23,"value":670}," 借助 ",{"type":18,"tag":26,"props":672,"children":675},{"href":673,"rel":674},"https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Fhooks",[30],[676],{"type":23,"value":677},"hooks",{"type":23,"value":679},"（如 ",{"type":18,"tag":256,"props":681,"children":683},{"className":682},[],[684],{"type":23,"value":685},"TeammateIdle",{"type":23,"value":687},"、",{"type":18,"tag":256,"props":689,"children":691},{"className":690},[],[692],{"type":23,"value":693},"TaskCompleted",{"type":23,"value":695},"），可以在你的检查通过之前，阻止队友收工或任务关闭。",{"type":18,"tag":19,"props":697,"children":698},{},[699],{"type":23,"value":700},"文档里的规模建议也很务实：从 3–5 个队友起步，每人 5–6 个任务，让每个队友独占一组文件；先从只读工作（评审、调研）开始，再尝试并行写代码。",{"type":18,"tag":50,"props":702,"children":704},{"id":703},"必须坦白的注意事项",[705],{"type":23,"value":703},{"type":18,"tag":19,"props":707,"children":708},{},[709],{"type":23,"value":710},"这是研究预览，行为也确实像研究预览：",{"type":18,"tag":325,"props":712,"children":713},{},[714,724,742,752,762],{"type":18,"tag":85,"props":715,"children":716},{},[717,722],{"type":18,"tag":37,"props":718,"children":719},{},[720],{"type":23,"value":721},"Token 消耗线性增长。",{"type":23,"value":723}," 每个队友都是独立的 Claude 实例，有自己的上下文窗口。五人团队的消耗大致相当于五个会话同时在跑。调研和评审场景通常值这个价，日常琐事则不值。",{"type":18,"tag":85,"props":725,"children":726},{},[727,732,734,740],{"type":18,"tag":37,"props":728,"children":729},{},[730],{"type":23,"value":731},"in-process 队友无法随会话恢复。",{"type":23,"value":733}," ",{"type":18,"tag":256,"props":735,"children":737},{"className":736},[],[738],{"type":23,"value":739},"\u002Fresume",{"type":23,"value":741}," 不会还原他们；负责人可能会给已不存在的队友发消息。",{"type":18,"tag":85,"props":743,"children":744},{},[745,750],{"type":18,"tag":37,"props":746,"children":747},{},[748],{"type":23,"value":749},"任务状态可能滞后。",{"type":23,"value":751}," 队友偶尔忘记把任务标记为完成，导致下游任务被阻塞，需要你去催一下。",{"type":18,"tag":85,"props":753,"children":754},{},[755,760],{"type":18,"tag":37,"props":756,"children":757},{},[758],{"type":23,"value":759},"结构性限制。",{"type":23,"value":761}," 每个会话只有一个团队；不能嵌套（队友不能再生成队友）；负责人在会话生命周期内固定，不能换人。",{"type":18,"tag":85,"props":763,"children":764},{},[765,770],{"type":18,"tag":37,"props":766,"children":767},{},[768],{"type":23,"value":769},"需要有人看着。",{"type":23,"value":771}," 文档说得很直白：让团队长时间无人值守地跑，浪费工作量的风险会上升。这是一个需要你掌舵的工具，不是一个可以定时调度的工具。",{"type":18,"tag":50,"props":773,"children":775},{"id":774},"它在ai-团队这条路上的位置",[776],{"type":23,"value":777},"它在\"AI 团队\"这条路上的位置",{"type":18,"tag":19,"props":779,"children":780},{},[781,783,789,791,796,798,804,806,812],{"type":23,"value":782},"Agent Teams 是今天体验多智能体协作最容易的方式：真实的并行、真实的点对点协调，而你始终在环内。如果这种工作方式对了你的胃口，接下来自然会问：一个更长期的 ",{"type":18,"tag":26,"props":784,"children":786},{"href":785},"\u002Fzh-CN\u002Fblog\u002Fwhat-is-an-ai-employee",[787],{"type":23,"value":788},"AI 员工",{"type":23,"value":790},"长什么样？如何把终端里的临时团队变成持久运转的团队？这就轮到",{"type":18,"tag":26,"props":792,"children":793},{"href":133},[794],{"type":23,"value":795},"生产级多智能体框架",{"type":23,"value":797},"和 ",{"type":18,"tag":26,"props":799,"children":801},{"href":800},"\u002Fzh-CN\u002Fblog\u002Fmcp-explained",[802],{"type":23,"value":803},"MCP",{"type":23,"value":805}," 这样的共享工具层出场了。想看把智能体组装成一个完整部门的全景图，可以从我们的 ",{"type":18,"tag":26,"props":807,"children":809},{"href":808},"\u002Fzh-CN\u002Fblog\u002Fhow-to-build-your-ai-team-2026",[810],{"type":23,"value":811},"2026 年 AI 团队搭建指南",{"type":23,"value":813},"读起。",{"type":18,"tag":815,"props":816,"children":817},"hr",{},[],{"type":18,"tag":19,"props":819,"children":820},{},[821,823,829],{"type":23,"value":822},"这些东西一发布，我们就会写。如果你也在搭建自己的 AI 团队，欢迎关注 BuildYour.Group，或者",{"type":18,"tag":26,"props":824,"children":826},{"href":825},"\u002F",[827],{"type":23,"value":828},"到首页加入早期体验候补名单",{"type":23,"value":830},"——我们想和你一起把它建起来。",{"type":18,"tag":832,"props":833,"children":834},"style",{},[835],{"type":23,"value":836},"html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}",{"title":7,"searchDepth":534,"depth":534,"links":838},[839,840,845,846,847,848,849,850],{"id":52,"depth":519,"text":55},{"id":141,"depth":519,"text":144,"children":841},[842,843,844],{"id":246,"depth":534,"text":249},{"id":266,"depth":534,"text":266},{"id":288,"depth":534,"text":288},{"id":306,"depth":519,"text":309},{"id":368,"depth":519,"text":368},{"id":440,"depth":519,"text":443},{"id":479,"depth":519,"text":479},{"id":703,"depth":519,"text":703},{"id":774,"depth":519,"text":777},"markdown","content:zh-CN:blog:claude-code-agent-teams-guide.md","content","zh-CN\u002Fblog\u002Fclaude-code-agent-teams-guide.md","zh-CN\u002Fblog\u002Fclaude-code-agent-teams-guide","md",1784513368742]