[{"data":1,"prerenderedAt":340},["ShallowReactive",2],{"article-how-i-use-multi-agent-collab\u002F":3},{"id":4,"title":5,"body":6,"cover":39,"date":301,"description":302,"draft":303,"extension":304,"faq":305,"featured":303,"meta":318,"navigation":319,"path":320,"readingTime":321,"seo":322,"seoKeywords":323,"stem":331,"summary":332,"tags":333,"type":337,"updated":338,"video":338,"__hash__":339},"articles\u002Farticles\u002Fhow-i-use-multi-agent-collab.md","看我如何用多 Agent 协同工作",{"type":7,"value":8,"toc":290},"minimark",[9,16,21,24,27,30,33,40,44,47,50,53,56,72,75,81,84,88,91,94,97,112,115,118,122,125,128,131,134,142,145,150,162,166,169,258,261,264,268,281,287],[10,11,12],"blockquote",{},[13,14,15],"p",{},"写小东西，一个助手就够。做 App 这种要好几个角色轮着上的活，我现在更愿意拆成一支小队，各聊各的。",[17,18,20],"h2",{"id":19},"我攒过不少用-agent-的方式","我攒过不少用 Agent 的方式",[13,22,23],{},"这一两年，常见用法我基本都试过。",[13,25,26],{},"有时就一个窗口聊到底。有时给同一个助手不停装 Skill、接工具，盼着它越来越全能。改脚本、改文案、查报错，这样用确实爽，少切换。",[13,28,29],{},"工作里另一些事，我会主动换招。",[13,31,32],{},"尤其做 App。节点多、角色杂，硬塞给一个助手，构建过程更容易飘。那种时候，我更喜欢多 Agent 一起干。",[13,34,35],{},[36,37],"img",{"alt":38,"src":39},"App 开发小队","\u002Fa\u002Fhow-i-use-multi-agent-collab\u002Fcover.jpg",[17,41,43],{"id":42},"做-app-时我更想要一支小队","做 App 时，我更想要一支小队",[13,45,46],{},"真做 App，很少是「你说一句，一个脑子答完」。",[13,48,49],{},"更像流水线。有人盯范围和优先级，有人写实现，有人找回归，有人做验收。节点不同，判断标准也不同。项目经理想的是里程碑和取舍，开发想接口和状态，测试想边界和复现，验收只关心这版能不能过。",[13,51,52],{},"这些全扔进同一个全能助手，聊天记录很快变成杂物间。上周的交互稿、某次接口命名吵架、半截失败日志、验收口径，缠成一团。下一句明明只是「登录失败的提示改清楚一点」，它却可能把测试用例或排期讨论一起拖出来。",[13,54,55],{},"所以我给自己定了默认班底：",[57,58,59,63,66,69],"ul",{},[60,61,62],"li",{},"项目经理：拆任务、排节点、盯范围",[60,64,65],{},"开发：实现、改代码",[60,67,68],{},"测试：找问题、写复现、盯回归",[60,70,71],{},"验收：对照目标，判断能不能过",[13,73,74],{},"美术资源真多的时候，再加一个美术 Agent。平时不加。小队够用就行，别为了显得很 AI 硬凑人数。",[76,77],"video-player",{"alt":78,"poster":79,"src":80},"自建 Agent 小队：添加专职角色后，由私人助理拉队友一起做 App","\u002Fa\u002Fhow-i-use-multi-agent-collab\u002Fagent-team-poster.jpg","\u002Fa\u002Fhow-i-use-multi-agent-collab\u002Fagent-team.mp4",[13,82,83],{},"多开几个窗口还不够。每个 Agent 尽量有自己的工作空间和技能，上下文隔开。开发不用背整份验收话术，测试也不必被半截产品讨论带跑。我自己用下来，这样做出的 App 更稳一点：谁负责什么清楚，返工也少扯皮。",[17,85,87],{"id":86},"什么时候多-agent-协同更划算","什么时候，多 Agent 协同更划算",[13,89,90],{},"多数人和我一样，日常还是单一助理用得多。单助手经常是对的，这点先说死。",[13,92,93],{},"任务同质、链路短、上下文本来就干净，硬拆小队只会多付切换成本。改句文案、问个 API、查一个孤立报错，一个助手通常更快。",[13,95,96],{},"我会认真考虑多 Agent，多半是下面几件事叠在一起：",[98,99,100,103,106,109],"ol",{},[60,101,102],{},"角色差得远，判断标准也不一样",[60,104,105],{},"上下文会互相污染，A 节点的细节混进 B，答案容易歪",[60,107,108],{},"需要交接或并行：开发改完测试接，测试踩坑再回到开发，别在同一段混聊里硬撑",[60,110,111],{},"想长期分桶：同类事下次还找同一个人，技能和材料留在它自己的空间里",[13,113,114],{},"做 App，这几条几乎都会撞上。写周报、翻译一段话、问天气，通常撞不上。",[13,116,117],{},"也有反例。你其实一个人在写原型，所有节点都经你手，拆四个 Agent 可能只是仪式感。小队有用，是因为角色真的分叉了，不是名词好听。",[17,119,121],{"id":120},"第三态一个入口调度多个实体干活","第三态：一个入口调度，多个实体干活",[13,123,124],{},"多 Agent 不等于你必须自己当总机，挨个点开联系人。",[13,126,127],{},"我更常这么用：私人助理当入口，背后多个专职实体干活。",[13,129,130],{},"你跟入口说目标，比如「这周把登录链路做到可验收」。它去拉开发、测试，必要时拉验收。真正动手的，还是各自带着干净上下文的专职 Agent。",[13,132,133],{},"这里有两件事别混：",[57,135,136,139],{},[60,137,138],{},"单入口：少点几下，省事",[60,140,141],{},"单上下文：所有角色记忆搅进同一锅",[13,143,144],{},"我要省事，不要搅锅。入口可以统一，上下文按实体分开。",[76,146],{"alt":147,"poster":148,"src":149},"私人助理作单入口，同时拉专职助手一起推进跨角色任务","\u002Fa\u002Fhow-i-use-multi-agent-collab\u002Fdispatch-hero-poster.jpg","\u002Fa\u002Fhow-i-use-multi-agent-collab\u002Fdispatch-hero.mp4",[13,151,152,153,161],{},"我在 ",[154,155,160],"a",{"href":156,"rel":157,"target":159},"https:\u002F\u002Fagentup.me",[158],"nofollow","_blank","AgentUp"," 里就是这么用的。通讯录里既有私人助理，也有专职角色；跨界时由入口调度，别把项目经理、开发、测试的脑子缝成一个。",[17,163,165],{"id":164},"单-agent-还是多-agent对照着选","单 Agent 还是多 Agent，对照着选",[13,167,168],{},"前面说散了，收成一张表更好查：",[170,171,172,188],"table",{},[173,174,175],"thead",{},[176,177,178,182,185],"tr",{},[179,180,181],"th",{},"看什么",[179,183,184],{},"更适合单 Agent 包办",[179,186,187],{},"更适合多 Agent 协同",[189,190,191,203,214,225,236,247],"tbody",{},[176,192,193,197,200],{},[194,195,196],"td",{},"任务形态",[194,198,199],{},"同质、短链路，一件事干完",[194,201,202],{},"多角色、多节点，标准不一样",[176,204,205,208,211],{},[194,206,207],{},"上下文",[194,209,210],{},"本来就干净，少历史包袱",[194,212,213],{},"容易互相污染，A 的细节会搅进 B",[176,215,216,219,222],{},[194,217,218],{},"协作方式",[194,220,221],{},"一个人连续做完就行",[194,223,224],{},"要交接或并行，比如开发完测试接",[176,226,227,230,233],{},[194,228,229],{},"记忆与技能",[194,231,232],{},"临时用用，不必长期分桶",[194,234,235],{},"同类事下次还找同一个人，技能要留住",[176,237,238,241,244],{},[194,239,240],{},"典型例子",[194,242,243],{},"改文案、问 API、查孤立报错",[194,245,246],{},"做 App：项目经理 \u002F 开发 \u002F 测试 \u002F 验收",[176,248,249,252,255],{},[194,250,251],{},"主要代价",[194,253,254],{},"角色一多就容易答偏、变脏",[194,256,257],{},"多付切换与调度成本；拆过头会很啰嗦",[13,259,260],{},"做 App，右边这列几乎都会撞上。写周报、翻译一段话、问天气，通常还是左边。懒得挨个点开时，可以私人助理当入口，专职实体在后面干活。",[13,262,263],{},"下次做 App，至少把「实现」和「测试」拆成两段干净上下文，看返工有没有少，就是最小验证。",[17,265,267],{"id":266},"我正在用-agentup-这样用","我正在用 AgentUp 这样用",[13,269,270,271,275,276,280],{},"这套选法不一定绑某个产品。我自己做进了 ",[154,272,160],{"href":273,"rel":274,"target":159},"https:\u002F\u002Fgithub.com\u002FMorvanZhou\u002Fagentup",[158],"：像加联系人一样加角色，私人助理可调度，专职 Agent 自带工作空间与技能。为什么做成通讯录，写在",[154,277,279],{"href":278},"\u002Fbuilds\u002Fagentup\u002F2026-10-22-why-agentup","这篇创造日志","里。",[13,282,283],{},[36,284],{"alt":285,"src":286},"AgentUp：像用微信那样用 AI","\u002Fa\u002Fhow-i-use-multi-agent-collab\u002Fagentup-banner.jpg",[13,288,289],{},"单一助理我还在用，而且用得很多。遇到 App 这种多身份、多节点的构建，我才会主动换成小队。隔开上下文，各干各的，我比较放心。",{"title":291,"searchDepth":292,"depth":292,"links":293},"",3,[294,296,297,298,299,300],{"id":19,"depth":295,"text":20},2,{"id":42,"depth":295,"text":43},{"id":86,"depth":295,"text":87},{"id":120,"depth":295,"text":121},{"id":164,"depth":295,"text":165},{"id":266,"depth":295,"text":267},"2026-10-22","单一助理用熟了之后，做 App 这类多角色、多节点的活，我更常换成项目经理、开发、测试、验收组成的多 Agent 小队。上下文隔离，专项专攻，构建更稳。",false,"md",[306,309,312,315],{"q":307,"a":308},"什么时候该用多 Agent 协同，而不是一个全能助手？","当任务要拆成异质角色、不同节点，且上下文容易互相污染时。比如做 App：立项与排期、写代码、测试、验收各管一段。角色越多、并行越深，一个全家桶助手越容易搅局。",{"q":310,"a":311},"多 Agent 团队一般怎么分工？","我做 App 时常用项目经理、开发、测试、验收。各自有工作空间和技能，上下文隔离。只有需要大量美术资源时，才再加一个美术 Agent，避免小队无故膨胀。",{"q":313,"a":314},"一定要自己点开每个 Agent 吗？","不必。务实做法是第三态：一个私人助理当单入口调度，背后多个专职实体干活。你跟入口说目标，由它拉对口角色；每个专职仍保持自己的上下文和技能边界。",{"q":316,"a":317},"多 Agent 协同可以怎么上手？","先判断要不要拆：同质短任务继续用单助手。做 App 这类多角色活，再按项目经理、开发、测试、验收分人，必要时加美术。平时可用私人助理当入口调度，专职 Agent 各自隔离上下文和技能；Skill 挂在对口角色上，别往同一个助手里乱堆。",{},true,"\u002Farticles\u002Fhow-i-use-multi-agent-collab",7,{"title":5,"description":302},[324,325,326,327,328,329,330],"多 Agent 协同","多 Agent 团队","单一 AI 助手","App 开发 Agent","Agent 上下文隔离","私人助理调度","多角色 Agent","articles\u002Fhow-i-use-multi-agent-collab","单一助理适合同质短任务；App 开发这类要项目经理、开发、测试、验收分节点推进的活，更适合多 Agent 协同。角色隔离上下文与技能，需要美术资源再加美术 Agent。最常见的务实形态是单入口调度加多专职实体。",[334,335,336],"AI 洞察","产品思考","AI 应用","article",null,"HaJ1Ve7FawgC4ndSa-lJvww6pZDJbyd6dSVgz-uX5gI",1793028248146]