[{"data":1,"prerenderedAt":1229},["ShallowReactive",2],{"build-logs-agentup":3,"build-log-agentup-":201},[4,203,322,481,762,1024],{"id":5,"title":6,"body":7,"date":164,"description":165,"draft":166,"extension":167,"faq":168,"meta":181,"navigation":182,"path":183,"project":184,"readingTime":185,"seo":186,"seoKeywords":187,"stem":195,"summary":196,"tags":197,"updated":201,"__hash__":202},"builds\u002Fbuilds\u002Fagentup\u002F2026-10-26-smoother-agent-teamwork.md","协作要丝滑：看得见谁在忙，交卷时还得带上文件",{"type":8,"value":9,"toc":155},"minimark",[10,18,26,29,32,36,39,42,45,48,54,57,60,63,66,69,75,78,81,84,140,143,146,149,152],[11,12,13],"p",{},[14,15],"img",{"alt":16,"src":17},"通讯录里需求、实现、质检三人，顶栏能看见谁还在协作","\u002Fb\u002Fagentup\u002Fcover-team-collab.webp",[11,19,20,25],{},[21,22,24],"a",{"href":23},"\u002Fbuilds\u002Fagentup\u002F2026-10-24-async-wait-and-continue\u002F","上一篇","把「能等、能接续」做出来了。真拿来用，别扭换了地方。",[11,27,28],{},"我让助理去问实现的同事。过一会儿它又问一遍，因为不记得刚才那单还没回来。侧栏上看不出谁正忙。专家终于回了，多半是一段话，该一起交的文件还在它自己的目录里。",[11,30,31],{},"人把活交给同事，会记得对方还在做，也不会把同一件事再派一次。交卷时常常是结论，外加文件。",[33,34,35],"h2",{"id":35},"现场要留在下一轮里",[11,37,38],{},"异步派出去之后，模型下一轮容易失忆，然后重复委托。",[11,40,41],{},"我把仍在进行的委托写进下一轮：执行中，或还在等你审批。没有进行中的就不写，免得每轮都占地方。最多 8 条，任务说明大约截到 120 字。写进去的话很直：任务重叠，或者还是同一个专家，先查进度，别再派一遍。",[11,43,44],{},"人这边也得看见。侧栏标出仍在异步协作里的 Agent。你打开的会话若正在被别人委托，会有占用提示。编排条补上已经跑了多久，大约半分钟刷新一次。",[11,46,47],{},"退出登录时，进行中的委托和异步任务一起停掉。人走了，后台不该还在转。",[11,49,50],{},[14,51],{"alt":52,"src":53},"当前会话显示被委托占用，侧栏里对应 Agent 仍在协作","\u002Fb\u002Fagentup\u002Fdelegation-occupancy.webp",[11,55,56],{},"重复派、看不清、退出后还在跑，都是用的时候撞上的。我没有先画一张架构图再往里填功能。",[33,58,59],{"id":59},"交卷不能只剩一段话",[11,61,62],{},"专家在自己的工作目录里写文件。父会话若只收到文字总结，文件还停在那边，交接就断了。",[11,64,65],{},"结束前加了一步：这次成品里，哪些要交给调用方。路径点清楚。没有要交的，也可以明确说空。有候选文件却还没点名，会再逼一轮，不默默全拷，也不默默不拷。确认之后才拷回父侧。",[11,67,68],{},"人交方案、补丁、清单，经常要带文件。Agent 之间我也想能带。",[11,70,71],{},[14,72],{"alt":73,"src":74},"专家点名回传的文件出现在调用方的成品列表里","\u002Fb\u002Fagentup\u002Fdeliver-artifacts.webp",[33,76,77],{"id":77},"先用开发三人组把小队跑起来",[11,79,80],{},"零散优化做着做着，我想要一个能反复用的编队，别每次临时拉人。",[11,82,83],{},"今天先落一支软件开发小队，三个人，交接顺序固定：",[85,86,87,103],"table",{},[88,89,90],"thead",{},[91,92,93,97,100],"tr",{},[94,95,96],"th",{},"谁",[94,98,99],{},"干什么",[94,101,102],{},"不干什么",[104,105,106,118,129],"tbody",{},[91,107,108,112,115],{},[109,110,111],"td",{},"需求",[109,113,114],{},"把范围问清楚，交出可执行说明",[109,116,117],{},"不写业务代码",[91,119,120,123,126],{},[109,121,122],{},"实现",[109,124,125],{},"按说明做功能",[109,127,128],{},"界面默认自己做，不另开一个人",[91,130,131,134,137],{},[109,132,133],{},"质检",[109,135,136],{},"对照说明做审查，给出复现和风险",[109,138,139],{},"不把需求再讲一遍",[11,141,142],{},"需求到实现，再到质检。中间跳两次。技能用现成开源包放进各自目录，并打成可安装的压缩包，免得每个角色从零手写。",[11,144,145],{},"编制我卡在大约三人。细岗并进技能，不按「再加一个专员」往上加。内容、学习、增长那些队还在名单里。开发队先做，因为最贴本机执行。后面还会在小队上加协作，这版没有做完。",[33,147,148],{"id":148},"收一下",[11,150,151],{},"我在补的是交接：下一轮记得谁还在忙，少重复派；交卷时可以点名文件；编队先用开发三人组，别在通讯录里堆一堆互不认识的角色。",[11,153,154],{},"现场提示、成品回传、这支小队，都是为了把活交给同事时顺一点。人数和界面以后还会改。",{"title":156,"searchDepth":157,"depth":157,"links":158},"",3,[159,161,162,163],{"id":35,"depth":160,"text":35},2,{"id":59,"depth":160,"text":59},{"id":77,"depth":160,"text":77},{"id":148,"depth":160,"text":148},"2026-10-26","用了一阵多 Agent，别扭变成重复派活、看不清谁在忙、交回来只有一段话。AgentUp 把进行中委托写进下一轮，并让专家点名回传文件；开发场景先落成三人小队。",false,"md",[169,172,175,178],{"q":170,"a":171},"AgentUp 怎么减少重复委托？","进行中的委托（执行中或等审批）会在下一轮对话里注入摘要，最多 8 条，任务说明截到约 120 字。同一专家、任务重叠时，先查进度，不要再派一遍。",{"q":173,"a":174},"委托结果为什么要单独回传文件？","交接经常同时给结论和附件。专家把文件写在自己的成品目录里，结束前要点名哪些属于这次任务、该交给调用方；也可以明确说没有要交的。确认后文件才会拷回父侧。",{"q":176,"a":177},"AgentUp 的开发小队现在是什么编制？","第一支落地的是软件产品开发队，三人：需求澄清与范围、功能实现、质检审查，按这个顺序交接。每队默认约三人，不按岗位无限加人。其他场景的小队还在后面。",{"q":179,"a":180},"用户怎么知道某个 Agent 正在被占用？","侧栏会标出仍在异步协作里的 Agent；你正打开的会话若在被委托干活，界面上有占用提示。编排条还会显示任务已进行多久。退出登录时，进行中的委托和异步任务会被停掉。",{},true,"\u002Fbuilds\u002Fagentup\u002F2026-10-26-smoother-agent-teamwork","agentup",5,{"title":6,"description":165},[188,189,190,191,192,193,194],"多 Agent 协作","Agent 团队管理","委托文件回传","异步任务现场感","Agent 小队分工","AgentUp 开发团队","减少重复委托","builds\u002Fagentup\u002F2026-10-26-smoother-agent-teamwork","Agent 协作要知道谁还在忙，交卷时也不只交文字。进行中委托会写进下一轮，避免重复派；专家要点名回传文件。开发场景先做成需求、实现、质检三人小队。",[198,199,200],"独立开发","产品思考","AI 应用",null,"uiJBxaw_38ZCBKNm90a-iz4jHax7csqFxJ88v5nkhY4",{"id":204,"title":205,"body":206,"date":292,"description":293,"draft":166,"extension":167,"faq":294,"meta":307,"navigation":182,"path":308,"project":184,"readingTime":185,"seo":309,"seoKeywords":310,"stem":318,"summary":319,"tags":320,"updated":201,"__hash__":321},"builds\u002Fbuilds\u002Fagentup\u002F2026-10-24-async-wait-and-continue.md","Agent 也该会等：别人还在干，自己先别卡死",{"type":8,"value":207,"toc":286},[208,214,220,223,227,230,233,236,240,247,250,253,259,262,266,269,272,275,278,280,283],[11,209,210],{},[14,211],{"alt":212,"src":213},"编排条同时显示专家委托与异步任务，可停止，下方提示正在整理","\u002Fb\u002Fagentup\u002Fcover-async-wake.webp",[11,215,216,219],{},[21,217,24],{"href":218},"\u002Fbuilds\u002Fagentup\u002F2026-10-23-delegation-fleet-visibility\u002F","说时间消耗型任务不该让整条对话干等；非专家长任务当时还在规划里。",[11,221,222],{},"这轮规划落地了。我想钉的是更具体的一点：Agent 也该会安排时间。可以等别人；别人交卷了，再干自己的后续。",[33,224,226],{"id":225},"干等不像协作","干等，不像协作",[11,228,229],{},"人把活交给同事，自己还能回邮件、改大纲。对方好了，再汇总。很少整段假死。",[11,231,232],{},"同步委托或阻塞式命令行做不到。助理一说「我去问设计」或「我去跑构建」，输入框就挂着转。你在等，它也不能干别的。",[11,234,235],{},"异步就是把这段等待从主回合上拿掉。先返回句柄，主会话继续；活做完再接。",[33,237,239],{"id":238},"专家和长命令进同一条编排","专家和长命令，进同一条编排",[11,241,242,243,246],{},"专家还是走委托，可一次派多名，默认异步。",[244,245],"br",{},"\n构建、监视、迁移、重脚本这类非专家活，走异步任务：发起、查询进度、必要时停止。社区后台任务包和当前运行时对不上，就自己对齐了同一套用法。",[11,248,249],{},"怎么选，写进工具说明和系统提示：大约二十秒内、本轮就要结果，可以同步等；更长、心里没底、还能干别的，就异步。命令行侧优先开异步任务，专家侧默认不等结果、或并行扇出。",[11,251,252],{},"顶栏也不再只显示委托。编排条里委托和异步任务排在一起，异步项用终端图标，都能停。同用户异步任务最多四个并发，多的排队，免得无限堆。",[11,254,255],{},[14,256],{"alt":257,"src":258},"编排条展开：一项委托查天气，一项异步任务在构建，均可停止","\u002Fb\u002Fagentup\u002Forchestration-bar.webp",[11,260,261],{},"主机环境也跟着走：Windows 主命令行是 PowerShell，mac \u002F Linux 是 bash；异步任务用同一套，审批也视同主命令行工具。",[33,263,265],{"id":264},"派出去之后还得接得上","派出去之后，还得接得上",[11,267,268],{},"只派不接，异步没意义。别人或后台完工了，谁来排下一步？",[11,270,271],{},"我们做了空闲唤醒。结果不先砸成一条可见杂讯；父会话空闲时，主动开一轮助手回复，把进展同步给你。编排条会进「正在整理…」，接续阶段可以看直播输出。你要是已经先发了下一条，就并进那一轮，不再另开一轮。",[11,273,274],{},"这样主助理才有空等专家和构建回来，再决定下一步，而不是永远停在「还在跑」。",[11,276,277],{},"取消仍是任务级。看不见、停不掉，我就不敢真并行。没有空闲唤醒，异步只是把假死挪到后台。",[33,279,148],{"id":148},[11,281,282],{},"我不在乎宣传「支持后台」。我在乎时间怎么用：能等，又不用整段卡死；别人完工了，还能接着排。",[11,284,285],{},"专家委托和异步任务共用编排与唤醒，就是为这个。数字和 UI 以后还会改，先按人怎么协作来。",{"title":156,"searchDepth":157,"depth":157,"links":287},[288,289,290,291],{"id":225,"depth":160,"text":226},{"id":238,"depth":160,"text":239},{"id":264,"depth":160,"text":265},{"id":148,"depth":160,"text":148},"2026-10-24","AgentUp 把专家委托和异步任务收进同一套编排：异步派出、顶栏可看可停，空闲后再开一轮接续。能等别人，也能在别人完工后接着干。",[295,298,301,304],{"q":296,"a":297},"AgentUp 的空闲唤醒是什么？","异步委托或异步任务结束后，若父会话正空闲，系统会主动开一轮助手回复，把结果同步给你；若你已经先发了下一条，则并进那一轮，不再另开一轮。",{"q":299,"a":300},"异步任务和委托专家有什么区别？","需要专职判断时委托给专家；长命令、构建、监视等非专家活走异步任务。两者都立刻返回编号，进同一条编排条，完成后走同一套空闲唤醒。",{"q":302,"a":303},"什么时候该异步，什么时候该同步等？","大约二十秒内、本轮就要用结果的，可以同步等待。更长、不确定、或还能干别的事，就异步返回任务编号，主会话先往下走。",{"q":305,"a":306},"异步任务跑着时用户能做什么？","继续聊天、看编排条进度、点开过程或点停止。编排条同时覆盖专家委托和异步任务，不必为两类耗时工作各学一套操作。",{},"\u002Fbuilds\u002Fagentup\u002F2026-10-24-async-wait-and-continue",{"title":205,"description":293},[311,312,313,314,315,316,317],"异步 Agent 协作","Agent 空闲唤醒","异步后台任务","多 Agent 编排","减少对话等待","AgentUp 编排条","subagent 异步","builds\u002Fagentup\u002F2026-10-24-async-wait-and-continue","Agent 也该会安排时间：能等别人，也能在别人完工后继续干。AgentUp 让专家委托与异步任务共用编排和空闲唤醒，时间消耗型工作不再绑架主对话。",[198,199,200],"v31ike3syqedtbxzj-tXfMyLJUYO0iekzFlTeSxudVg",{"id":323,"title":324,"body":325,"date":292,"description":452,"draft":166,"extension":167,"faq":453,"meta":466,"navigation":182,"path":467,"project":184,"readingTime":185,"seo":468,"seoKeywords":469,"stem":477,"summary":478,"tags":479,"updated":201,"__hash__":480},"builds\u002Fbuilds\u002Fagentup\u002F2026-10-24-context-mount-and-discovery.md","上下文要独立，感知能扩，能力还得找得到",{"type":8,"value":326,"toc":446},[327,333,340,343,346,349,352,364,371,377,380,384,387,394,401,404,408,411,428,431,434,440,443],[11,328,329],{},[14,330],{"alt":331,"src":332},"工作区「项目」页签下列出已挂载的本机目录，左侧为多 Agent 通讯录","\u002Fb\u002Fagentup\u002Fcover-context-mount.webp",[11,334,335,339],{},[21,336,338],{"href":337},"\u002Fbuilds\u002Fagentup\u002F2026-10-22-why-agentup\u002F","初衷","是分人办事：专职 Agent，上下文分开，别往同一个全能助手里堆。",[11,341,342],{},"小队一旦要变大，目录、发现、装配会一起冒出来。这轮主要啃这几块。",[33,344,345],{"id":345},"工作目录也算上下文",[11,347,348],{},"我对 Agent 的定义是互相独立的上下文。聊天记录分开还不够，工作目录也是它能感知到的范围。",[11,350,351],{},"每个 Agent 可以有自己的工作目录，默认别互相踩。开发改代码、测试盯复现、资料助手整理文献，磁盘上也少搅在一起。",[11,353,354,355,359,360,363],{},"可感知范围不该锁死在某一个路径。真干活时常要碰到用户本机上的项目盘。所以工作区加了项目根：在「项目」Tab 挂本机目录，写进该次访问的 ",[356,357,358],"code",{},"project_roots","；文件工具只允许在 ",[356,361,362],{},"visitCwd + project_roots"," 里动。",[11,365,366,367,370],{},"越界不默默放行。审批会提议挂载（",[356,368,369],{},"project_mount","）；你点头后写入再继续。once 只管这一次；session \u002F agent 可以记路径允许，后面同类越界少点几次。敏感路径照样拒。",[11,372,373],{},[14,374],{"alt":375,"src":376},"越界访问时弹出挂载审批：可拒绝，或允许挂载本机项目目录","\u002Fb\u002Fagentup\u002Fproject-mount-approval.webp",[11,378,379],{},"独立是为了别干扰。能挂载，是需要时把感知撑开。不是一上来把整块硬盘交给同一个 Agent。",[33,381,383],{"id":382},"人一多就得会找","人一多，就得会找",[11,385,386],{},"Agent 多了，工具多了，最省事是每轮把花名册和工具表全喂进去。人少还凑合；人一多，上下文先炸。",[11,388,389,390,393],{},"所以发现改成检索。不再每轮塞「小花名册全量」。按用户话术打分，有命中才注入 Top-K；模型也可以主动 ",[356,391,392],{},"agent_search","。协作偏好里写着：委托前先搜再派。",[11,395,396,397,400],{},"MCP 也一样：默认延迟暴露，先 ",[356,398,399],{},"tool_search"," 再调用，别假设工具已经全声明在上下文里。",[11,402,403],{},"实体可以很多。当前这一轮看见的，得少而准。发现不是多一个搜索框，是挡上下文爆炸的闸。",[33,405,407],{"id":406},"配起来太麻烦app-就活不了","配起来太麻烦，app 就活不了",[11,409,410],{},"只有隔离和发现还不够。每个 Agent 都要从零手搓技能的话，通讯录里永远就那几个示范角色。",[11,412,413,414,417,418,417,421,417,424,427],{},"这轮接上社区 Skill 的搜索和安装，包校验也放宽一点，兼容更多 frontmatter。系统提示里收紧工具选用：能用 ",[356,415,416],{},"grep"," \u002F ",[356,419,420],{},"find",[356,422,423],{},"read",[356,425,426],{},"edit"," 就别滥用 shell，少一次审批，也少一次跑偏。",[11,429,430],{},"我想降低装配成本，让里面的 Agent 真的多起来。多起来之后，独立目录和发现才用得上；否则架子搭了，屋里没人。",[33,432,433],{"id":433},"和时间那条怎么拼",[11,435,436,439],{},[21,437,438],{"href":23},"异步编排","管等待和接续。这篇管看见什么、找得到谁、装得上什么。",[11,441,442],{},"别互相干扰，靠独立 cwd 和会话。偶尔要更大感知，就挂项目根，越界先问你。实体和工具变多，靠检索，不靠全量塞提示。app 要活，还得 Skill 好配。",[11,444,445],{},"小队要长大，时间和空间这两边都得有。",{"title":156,"searchDepth":157,"depth":157,"links":447},[448,449,450,451],{"id":345,"depth":160,"text":345},{"id":382,"depth":160,"text":383},{"id":406,"depth":160,"text":407},{"id":433,"depth":160,"text":433},"AgentUp 把工作目录当成上下文的一部分：默认隔离，又能动态挂载项目根。team 变大后用检索发现 Agent 和工具，再用社区 Skill 降低装配成本，让 app 里的 Agent 真正多起来。",[454,457,460,463],{"q":455,"a":456},"为什么 Agent 的工作目录也算上下文？","独立上下文不只是聊天记录分开。能读哪些路径、改哪些文件，直接决定它看见什么。默认各自 cwd，减少互相干扰；需要时再挂载额外项目根，扩大感知范围。",{"q":458,"a":459},"动态挂载项目目录安全吗？","文件工具默认只在 visit cwd 与已挂载的 project_roots 内。越界访问会走审批，提议挂载目录；你允许后才写入挂载并继续。敏感路径仍会拒绝。",{"q":461,"a":462},"Agent 和工具变多时，AgentUp 怎么避免上下文爆炸？","不再每轮塞小花名册全量。按话术检索，有命中才注入 Top-K；模型也可主动 agent_search。MCP 工具默认延迟暴露，先 tool_search 再调用，和 Agent 发现同一思路。",{"q":464,"a":465},"社区 Skill 解决什么问题？","让 Agent 好配、好添能力。可从社区源搜索安装 Skill，包校验也兼容更多 frontmatter，降低从零手写配置的成本，app 里才容易真的「Agent 多起来」。",{},"\u002Fbuilds\u002Fagentup\u002F2026-10-24-context-mount-and-discovery",{"title":324,"description":452},[470,471,472,473,474,475,476],"Agent 工作目录","项目目录挂载","Agent 发现机制","减少上下文爆炸","社区 Skill 安装","多 Agent 上下文隔离","AgentUp 工作区","builds\u002Fagentup\u002F2026-10-24-context-mount-and-discovery","Agent 互相独立，工作目录也是上下文边界。动态挂载扩大感知，检索发现避免全量塞提示，社区 Skill 降低配置难度，小队才能变多而不炸。",[198,199,200],"AFWROZNIs6EwYvTcdAmV8yJmI-D6XqHZXhexkCS2NBk",{"id":482,"title":483,"body":484,"date":731,"description":732,"draft":166,"extension":167,"faq":733,"meta":746,"navigation":182,"path":747,"project":184,"readingTime":748,"seo":749,"seoKeywords":750,"stem":758,"summary":759,"tags":760,"updated":201,"__hash__":761},"builds\u002Fbuilds\u002Fagentup\u002F2026-10-23-delegation-fleet-visibility.md","时间消耗型任务，不该让整条对话干等",{"type":8,"value":485,"toc":724},[486,492,498,501,504,512,516,519,522,525,540,543,562,565,569,572,586,589,592,596,599,602,605,644,647,653,657,660,667,673,684,687,690,696,700,703,712,715],[11,487,488],{},[14,489],{"alt":490,"src":491},"对话顶栏显示多项异步委托进行中，列表里可点停止","\u002Fb\u002Fagentup\u002Fcover-delegation-fleet.webp",[11,493,494,497],{},[21,495,496],{"href":337},"初衷文","说的是关系：通讯录里的专职 Agent，跨界再组队。",[11,499,500],{},"组队一旦真跑起来，下一刀切在时间上。",[11,502,503],{},"查文献、估交互、跑构建、盯一段长日志，这些都是时间消耗型。若主会话必须同步等完，用户体感就是：助理说「我去办」，然后整条对话假死。",[11,505,506,507,511],{},"最近 coding agent 里很扎眼的一个模式，就是把这类活",[508,509,510],"strong",{},"异步化","：先返回句柄，主会话继续聊；做完再回注。AgentUp 这轮，是把这个模式落到「专家 subagent」上，并准备把普通长任务也收进同一套管法。",[33,513,515],{"id":514},"同步等结果只适合短活","同步等结果，只适合短活",[11,517,518],{},"最早的委托很朴素：指定专家，写有界 brief，默认同步等回来。",[11,520,521],{},"一两分钟能结束的活，这样最干净，结果直接进本轮 tool result。",[11,523,524],{},"一拉长就不行。你要同时查结构、估交互、扫风险，串行同步等于把等待时间加总。主助理也不能边等边回你别的问题。",[11,526,527,528,531,532,535,536,539],{},"所以工具改成可扇出：",[356,529,530],{},"tasks=[{agent_id, brief}, ...]","。多项默认异步，立刻返回各 ",[356,533,534],{},"task_id","；这轮就必须汇总，再显式 ",[356,537,538],{},"wait=true","。",[11,541,542],{},"边界写死：",[544,545,546,550,553],"ul",{},[547,548,549],"li",{},"一次最多扇出 8 个专家子任务",[547,551,552],{},"同用户并发闸门 4，超出排队",[547,554,555,558,559,561],{},[356,556,557],{},"check_delegation"," 可按多个 ",[356,560,534],{}," 汇合",[11,563,564],{},"异步不是「扔出去不管」。是把阻塞从主 turn 上拿掉，换成可查询的任务。",[33,566,568],{"id":567},"业界也在把-subagent-往后台推","业界也在把 subagent 往后台推",[11,570,571],{},"这不是我一个人的偏好。开源 agent 框架和主流 coding agent 文档里，越来越常见同一套分工：",[544,573,574,580],{},[547,575,576,579],{},[508,577,578],{},"同步 subagent","：主会话阻塞到子任务结束。适合短、确定、下一步立刻依赖结果的活。",[547,581,582,585],{},[508,583,584],{},"异步 \u002F background subagent","：立刻返回 task id；可检查、可改指令、可取消；完成后用通知或新一轮回注，而不是卡死当前 tool 调用。",[11,587,588],{},"有的实现还强调：结果要等主会话空闲再注入，避免插在 tool result 和 assistant 消息中间，把角色交替和缓存搅乱。AgentUp 的异步委托也走「完成后空闲唤醒」，同一类顾虑。",[11,590,591],{},"一句话：时间消耗型工作，默认不该绑架主对话的时钟。",[33,593,595],{"id":594},"专家异步了普通长任务也要","专家异步了，普通长任务也要",[11,597,598],{},"耗时的不只是「问另一个 Agent」。",[11,600,601],{},"长命令、本地构建、盯日志、跑测试，往往根本不需要招募专家，但一样能把对话拖死。常见做法是开异步任务：跑起来拿编号，可查进度，不行就停。",[11,603,604],{},"我想要的形状是：",[85,606,607,620],{},[88,608,609],{},[91,610,611,614,617],{},[94,612,613],{},"类型",[94,615,616],{},"谁干",[94,618,619],{},"怎么异步",[104,621,622,633],{},[91,623,624,627,630],{},[109,625,626],{},"专家子代理",[109,628,629],{},"已招募的专职 Agent",[109,631,632],{},"委托工具，扇出默认不等待",[91,634,635,638,641],{},[109,636,637],{},"非专家长任务",[109,639,640],{},"本机异步任务",[109,642,643],{},"规划中与委托共用编排条",[11,645,646],{},"今天落地的是左边：专家委托异步、扇出、取消、Fleet。右边在架构里写明了：自研对齐后台作业契约，Desktop Fleet 扩展成「委托 ∪ 后台作业」统一条。第三方包和当前运行时版本对不上，不能直接装，所以不硬接。",[11,648,649,650],{},"目标一致：",[508,651,652],{},"凡是会烧时间的，都先变成可监视的后台项，主会话先往下走。",[33,654,656],{"id":655},"异步之后人要看得见停得掉","异步之后，人要看得见、停得掉",[11,658,659],{},"派出去减等待，只对了一半。看不见、停不掉，用户还是不敢放开用。",[11,661,662,663,666],{},"子过程不塞进父消息 parts，免得主对话被过程日志淹没。顶栏加 ",[508,664,665],{},"Fleet","：轮询进行中的异步项，显示头像、短 brief、数量；点开看过程；列表可展开。",[11,668,669],{},[14,670],{"alt":671,"src":672},"Fleet 展开条：两位专家在跑，每行都有停止按钮","\u002Fb\u002Fagentup\u002Ffleet-bar.webp",[11,674,675,676,679,680,683],{},"取消是任务级的：",[356,677,678],{},"cancel_delegation","、HTTP cancel、Fleet \u002F 过程弹窗上的「停止」。任务 AbortController 和父 turn signal 并联；终态 ",[356,681,682],{},"cancelled"," 再回注。这跟「停止当前回复」不是一层：后者打断眼前这一轮，前者打断已经在后台转的活。",[11,685,686],{},"审批仍要过人。委托范围可记到 session 或 agent。异步放大了并行，也放大了误派成本；许可和中止得一起给。",[11,688,689],{},"消息侧的 Artifacts 则补另一面：异步跑完常会改文件，列表里能看见新增 \u002F 修改 \u002F 删除，不只读一段总结。",[11,691,692],{},[14,693],{"alt":694,"src":695},"消息「产物」页签列出新增、修改、删除的文件路径","\u002Fb\u002Fagentup\u002Fartifacts-panel.webp",[33,697,699],{"id":698},"我在优化的等待到底是什么","我在优化的等待，到底是什么",[11,701,702],{},"不是让模型「算得更快」。是让用户少坐在假死的输入框前面。",[11,704,705,706,708,709,711],{},"短活：同步等，结果马上用。",[244,707],{},"\n长活、可并行活：异步派，主会话继续聊，Fleet 盯着，跑偏就停。",[244,710],{},"\n专家和非专家：两条入口，一套可见与取消。",[11,713,714],{},"扇出 8、并发 4 只是本机默认刹车。后台作业统一进 Fleet，还在后面。数字和 UI 都会再调；模式先钉住。",[11,716,717,718,539],{},"想试用：",[21,719,720],{"href":720,"rel":721,"target":723},"https:\u002F\u002Fagentup.me",[722],"nofollow","_blank",{"title":156,"searchDepth":157,"depth":157,"links":725},[726,727,728,729,730],{"id":514,"depth":160,"text":515},{"id":567,"depth":160,"text":568},{"id":594,"depth":160,"text":595},{"id":655,"depth":160,"text":656},{"id":698,"depth":160,"text":699},"2026-10-23","Coding agent 正在把 subagent 和长任务异步化：先返回 task id，主会话继续聊，做完再回注。AgentUp 把专家委托默认异步，并规划把构建、长 shell 等非专家任务收进同一套 Fleet，减少等待。",[734,737,740,743],{"q":735,"a":736},"什么是 coding agent 里的异步任务模式？","主 Agent 派发子任务后立刻拿到 task id，不等结果返回就继续和用户交互；子任务在后台跑，完成后再通知或回注。适合长跑、可并行、或中途可能要改方向的工作。",{"q":738,"a":739},"AgentUp 的异步委托和同步委托差在哪？","同步（wait=true）把结果当本轮 tool result，主会话干等。异步（wait=false，扇出默认）立刻返回 task_id，专家在后台跑，完成后空闲唤醒父会话；顶栏 Fleet 可看进度、可停止。",{"q":741,"a":742},"AgentUp 为什么要把 subagent 和普通任务都异步化？","耗时的不只是「问专家」。长 shell、构建、监视同样会拖死对话。专家走 delegate；非专家长任务走后台作业。两边共用可见与取消模型，主会话才不用为每类耗时工作单独干等。",{"q":744,"a":745},"异步派出去之后怎么管？","桌面顶栏 Fleet 列出进行中的异步项；可点开过程，可点「停止」。工具侧有 check_delegation \u002F cancel_delegation。扇出最多 8，同用户并发闸门 4，避免后台任务无限堆叠。",{},"\u002Fbuilds\u002Fagentup\u002F2026-10-23-delegation-fleet-visibility",6,{"title":483,"description":732},[751,752,753,754,755,756,757],"异步 Agent 任务","coding agent 异步","subagent 后台执行","多 Agent 委托","Agent 任务取消","减少 Agent 等待","AgentUp Fleet","builds\u002Fagentup\u002F2026-10-23-delegation-fleet-visibility","同步委托会把整轮对话卡在时间消耗型工作上。AgentUp 把专家 subagent 默认异步扇出，用 Fleet 监视与取消；非专家长任务也按同一套后台作业模型规划，让主会话先往下走。",[198,199,200],"xaVux_FAgcv0jKv61AZQjnoiyyjPs1bNgD03Vl5stek",{"id":763,"title":764,"body":765,"date":731,"description":995,"draft":166,"extension":167,"faq":996,"meta":1009,"navigation":182,"path":1010,"project":184,"readingTime":748,"seo":1011,"seoKeywords":1012,"stem":1020,"summary":1021,"tags":1022,"updated":201,"__hash__":1023},"builds\u002Fbuilds\u002Fagentup\u002F2026-10-23-desktop-windows-shipping.md","安装慢往往不是包太大，是小文件太多",{"type":8,"value":766,"toc":989},[767,773,778,781,784,788,794,805,812,827,833,837,843,876,882,903,906,910,917,932,946,949,952,959,962,973,980,983],[11,768,769],{},[14,770],{"alt":771,"src":772},"官网下载页提供 macOS 与 Windows 入口，版本 v0.1.1","\u002Fb\u002Fagentup\u002Fcover-desktop-download.webp",[11,774,775,777],{},[21,776,496],{"href":337},"之后，桌面端要从「我自己能跑」走到「别人装得上」。",[11,779,780],{},"我更在乎的不是下载按钮画在哪，而是打包规则本身：",[11,782,783],{},"包怎么尽量小？Windows 安装为什么拖成泥？这两件事经常被混成一句「优化安装包」。实际不是一回事。",[33,785,787],{"id":786},"慢的经常是文件个数不是总大小","慢的经常是文件个数，不是总大小",[11,789,790,791,539],{},"本机 Agent 带着一整棵 runtime：会话、工具、委托都在那一层。依赖是真·",[356,792,793],{},"node_modules",[11,795,796,797,800,801,804],{},"Electron 默认会把渲染层打进 ",[356,798,799],{},"app.asar","。runtime 不行。它用 ",[356,802,803],{},"ELECTRON_RUN_AS_NODE=1"," 起子进程，没有 asar 虚拟文件系统，依赖必须落在真实路径上。",[11,806,807,808,811],{},"于是慢点就暴露了：松散的 ",[356,809,810],{},"resources\u002Fruntime\u002Fnode_modules"," 里是上万小文件。NSIS 逐条写出时极慢；DMG \u002F zip 拷贝也慢。你盯安装包 MB 数，以为还行；用户盯进度条，已经等疯了。",[11,813,814,815,818,819,822,823,826],{},"对策很土，但有效：物化之后，立刻打成单个 ",[356,816,817],{},"node_modules.tgz","，再删掉松散树。安装器只处理一个文件；展开交给系统 ",[356,820,821],{},"tar","。Windows 在 NSIS 的 ",[356,824,825],{},"customInstall"," 里解；mac 没有安装脚本，首次启动再解一次，大约十几秒，之后跳过。",[11,828,829,832],{},[356,830,831],{},".tgz"," 本身已是 gzip。还得告诉 NSIS 别再压一遍，否则白费 CPU，安装也不更快。",[33,834,836],{"id":835},"打包规则写成流水线不靠临场发挥","打包规则写成流水线，不靠临场发挥",[11,838,839,842],{},[356,840,841],{},"afterPack"," 里顺序是死的：",[844,845,846,860,870],"ol",{},[547,847,848,851,852,855,856,859],{},[508,849,850],{},"prune","：砍文档、sourcemap、",[356,853,854],{},"@types","、常见 changelog \u002F readme。只删确定用不到的；不删业务依赖，也不误伤包内部叫 ",[356,857,858],{},"doc"," 的目录。",[547,861,862,865,866,869],{},[508,863,864],{},"materialize","：把构建机上的绝对路径 symlink 改成包内相对链接，保留「一份 ",[356,867,868],{},".pnpm"," 实体」。逃逸出包的链接直接报错，不静默出坏包。",[547,871,872,875],{},[508,873,874],{},"pack tgz","：打归档，删松散树。",[11,877,878],{},[14,879],{"alt":880,"src":881},"物化：绝对路径 symlink 改成包内相对链接，体积仍约一份依赖","\u002Fb\u002Fagentup\u002Fruntime-materialize.webp",[11,883,884,885,887,888,891,892,895,896,899,900,902],{},"为什么不把整棵 ",[356,886,793],{}," 写进 ",[356,889,890],{},"extraResources","？electron-builder 会按忽略规则跳过它。所以 ",[356,893,894],{},"package.json"," 只带 ",[356,897,898],{},"dist"," 和 ",[356,901,894],{},"，依赖在 afterPack 注入。",[11,904,905],{},"这三条合在一起，才叫「打包规则」。少一步，要么体积炸，要么装到别人机器起不来，要么安装器卡在小文件上。",[33,907,909],{"id":908},"mac-保体积windows-保能解开","mac 保体积，Windows 保能解开",[11,911,912,913,916],{},"同一份依赖，打 tgz 时 ",[356,914,915],{},"follow"," 必须按平台分叉。",[11,918,919,920,923,924,927,928,931],{},"mac：",[356,921,922],{},"follow=false","，保留相对 symlink。dmg 大概在 ",[508,925,926],{},"160MB"," 量级。我踩过坑：darwin 误开 follow，链接被解成多份实体，dmg 胀到 ",[508,929,930],{},"216MB+","。体积回归时，先查是不是跟错了平台开关。",[11,933,934,935,938,939,941,942,945],{},"Windows：",[356,936,937],{},"follow=true","，必须解成实体文件。Windows 自带的 tar 要求创建 symlink 时目标已经存在；归档里顶层链接常常先于 ",[356,940,868],{}," 实体出现，会大片 ",[356,943,944],{},"Cannot create symlink","。setup 会更大，但 NSIS 仍然只解压一个 tgz，安装速度还能保住。",[11,947,948],{},"两边目标不一样：mac 优先「一份依赖」；win 优先「解得开」。用同一套 follow，两边都会倒霉。",[33,950,951],{"id":951},"我实际在优化什么",[11,953,954,955,958],{},"下载页、Windows 构建链、版本提到 ",[356,956,957],{},"0.1.1","，同日都做了。那是入口。",[11,960,961],{},"这篇文章想记清楚的是入口后面的判断：",[544,963,964,967,970],{},[547,965,966],{},"包要小：先 prune，再靠相对链接别复制多份。",[547,968,969],{},"安装要快：别让安装器面对上万小文件，打成单个 tgz。",[547,971,972],{},"规则要分平台：mac 禁 follow，win 必须 follow；注释和测试里都写死了回归信号。",[11,974,975,976,979],{},"同日产品侧还有",[21,977,978],{"href":218},"委托舰队与取消","。协作体验和安装体验是两件事；装不上，前面都白做。",[11,981,982],{},"签名、杀软、不同 Windows 版本差异后面还会冒。当前这套先把「体积」和「小文件拖垮安装」这两条红线钉住。",[11,984,717,985,539],{},[21,986,987],{"href":987,"rel":988,"target":723},"https:\u002F\u002Fagentup.me\u002Fdownload",[722],{"title":156,"searchDepth":157,"depth":157,"links":990},[991,992,993,994],{"id":786,"depth":160,"text":787},{"id":835,"depth":160,"text":836},{"id":908,"depth":160,"text":909},{"id":951,"depth":160,"text":951},"AgentUp 桌面打包先 prune 再物化，最后把 runtime 的 node_modules 打成单个 tgz。mac 保相对链接控体积，Windows 解引用换可装；NSIS 只处理一个文件，安装才不拖成泥潭。",[997,1000,1003,1006],{"q":998,"a":999},"为什么 Electron 应用在 Windows 上安装会特别慢？","常见原因不是安装包总字节数，而是安装器要逐个写出上万个小文件。NSIS 对松散 node_modules 极慢。AgentUp 把 runtime 依赖打成单个 node_modules.tgz，安装期用系统 tar 一次展开。",{"q":1001,"a":1002},"AgentUp 怎么控制桌面安装包体积？","打包前 prune 掉文档、sourcemap、@types 等确定用不到的内容；materialize 保留 .pnpm 一份实体加相对 symlink；mac 打 tgz 时 follow=false，避免把链接解成多份拷贝把 dmg 胀大。",{"q":1004,"a":1005},"为什么 mac 和 Windows 的打包规则不一样？","mac 用相对 symlink 保体积，dmg 大约 160MB 量级；Windows 的 tar 不能在目标尚未落地时创建 symlink，所以 win 打包必须 follow 成实体文件。setup 会更大，但安装器仍只解压一个 tgz，安装速度还能保住。",{"q":1007,"a":1008},"为什么 runtime 的 node_modules 不放进 app.asar？","runtime 子进程以 ELECTRON_RUN_AS_NODE 跑，拿不到 asar 虚拟文件系统，依赖必须是真实路径。所以 extraResources 只放 dist 和 package.json，node_modules 在 afterPack 里物化后再打成 tgz 注入。",{},"\u002Fbuilds\u002Fagentup\u002F2026-10-23-desktop-windows-shipping",{"title":764,"description":995},[1013,1014,1015,1016,1017,1018,1019],"Electron 打包体积优化","Windows NSIS 安装慢","node_modules tgz","pnpm symlink 打包","runtime 依赖裁剪","桌面应用打包规则","AgentUp 桌面端","builds\u002Fagentup\u002F2026-10-23-desktop-windows-shipping","本机 Agent 的 runtime 不能塞进 asar。AgentUp 用 prune、物化、再打成单个 node_modules.tgz 控体积；并针对 Windows「上万小文件拖垮 NSIS」把安装器改成只解压一个归档。",[198,199],"lByZpJ_KXrCGD3-3-axcLl1dc8VRZXD5P_lCZws-t58",{"id":1025,"title":1026,"body":1027,"date":1199,"description":1200,"draft":166,"extension":167,"faq":1201,"meta":1214,"navigation":182,"path":1215,"project":184,"readingTime":185,"seo":1216,"seoKeywords":1217,"stem":1225,"summary":1226,"tags":1227,"updated":201,"__hash__":1228},"builds\u002Fbuilds\u002Fagentup\u002F2026-10-22-why-agentup.md","私人助理越装越臃肿，我想要的是一支 Agent 小队",{"type":8,"value":1028,"toc":1192},[1029,1035,1039,1042,1045,1052,1055,1059,1062,1069,1072,1083,1089,1095,1099,1110,1113,1121,1124,1138,1144,1150,1153,1157,1160,1166,1169,1172,1175,1178,1181,1184,1187,1190],[11,1030,1031],{},[14,1032],{"alt":1033,"src":1034},"自建小队：产品、设计、前端、后端各管一块，也能互相协作","\u002Fb\u002Fagentup\u002Ffeature-agent-team.png",[33,1036,1038],{"id":1037},"一个助手装到后来什么都答不准","一个助手，装到后来什么都答不准",[11,1040,1041],{},"最近用个人助手型的 Agent app，有一种熟悉的别扭。",[11,1043,1044],{},"你给它装论文、装代码、装日程、装天气。一开始很爽，什么都能问。过一阵子，你只想知道「明天出差要不要带伞」，它却还拖着上周的论文讨论、某次报错日志、半截周报提纲。",[11,1046,1047,1048,1051],{},"不是模型突然变笨了。是",[508,1049,1050],{},"这一个助手被喂得太杂","，窄任务也得背全家桶。",[11,1053,1054],{},"人办事不是这样的。微信里你找谁办什么事，对方有自己的背景，上下文天然隔离。我想要的，越来越像后者。",[33,1056,1058],{"id":1057},"agents-正在变专也适合组队","Agents 正在变专，也适合组队",[11,1060,1061],{},"过去一年，我越来越明显地感到：Agent 不再只适合做成「我的全能私人助理」。",[11,1063,1064,1065,1068],{},"它们在往",[508,1066,1067],{},"专业化、深度化","走。有人擅长读文献，有人擅长盯天气，有人擅长把桌面文件整理成交付物。单看每一个，边界清楚；合在一起，才像一支能协作的小队。",[11,1070,1071],{},"市面上常见的三条路，各自有价值，但都没接到我想要的体验：",[544,1073,1074,1077,1080],{},[547,1075,1076],{},"把 Bot 挂在人机聊天产品上：主体仍是人，Bot 是受限插件。",[547,1078,1079],{},"模板或技能商店：你拿到的是私人副本，再往同一个实例里继续装。",[547,1081,1082],{},"手机遥控「我电脑上的那个助手」：本地执行有了，公共实体和分人办事还是缺。",[11,1084,1085,1086],{},"我想做的不是再堆一个更全能的私人助理，而是换一种关系：",[508,1087,1088],{},"通讯录里有一群专职 Agent，跨界时再组队。",[11,1090,1091],{},[14,1092],{"alt":1093,"src":1094},"像加好友一样找到天气助手，开聊就能问会不会下雨","\u002Fb\u002Fagentup\u002Ffeature-add-assistant.png",[33,1096,1098],{"id":1097},"所以做了-agentup","所以做了 AgentUp",[11,1100,1101,1102,1105,1106,539],{},"产品叫 AgentUp。想试用或下载，去官网：",[21,1103,720],{"href":720,"rel":1104,"target":723},[722],"。源码开源在 GitHub：",[21,1107,1108],{"href":1108,"rel":1109,"target":723},"https:\u002F\u002Fgithub.com\u002FMorvanZhou\u002Fagentup",[722],[11,1111,1112],{},"它要回答的不是「怎么再写一个 coding agent」，而是：",[1114,1115,1116],"blockquote",{},[11,1117,1118],{},[508,1119,1120],{},"像用微信那样用 AI。",[11,1122,1123],{},"具体一点：",[544,1125,1126,1129,1132,1135],{},[547,1127,1128],{},"像加好友一样，把助手加进通讯录",[547,1130,1131],{},"一个助手只聊一类事，不胡搅在一起",[547,1133,1134],{},"事情跨界了，再去拉别的助手一起帮你",[547,1136,1137],{},"能在你电脑里看文件、把结果写回去；聊天默认留在本机",[11,1139,1140],{},[14,1141],{"alt":1142,"src":1143},"通讯录里有私人助理和专职帮手；跨界问题由助理拉队友一起查","\u002Fb\u002Fagentup\u002Ffeature-hero.png",[11,1145,1146,1147],{},"一句话切分也写在产品边界里：",[508,1148,1149],{},"编码工作台管工具；AgentUp 承载服务关系。",[11,1151,1152],{},"小队可以自己组。每人有专长，跨领域时再互相拉一把，而不是把所有本事塞回同一个人。",[11,1154,1155],{},[14,1156],{"alt":1033,"src":1034},[11,1158,1159],{},"干活也尽量留在你这边：读桌面文件，把结果写回去；聊天默认存在本机。",[11,1161,1162],{},[14,1163],{"alt":1164,"src":1165},"助手在本机读桌面文件，再把 PPT 等结果写回电脑","\u002Fb\u002Fagentup\u002Ffeature-on-computer.png",[33,1167,1168],{"id":1168},"我刻意不碰的几条路",[11,1170,1171],{},"初衷文先把边界说清，后面才不至于做成「又一个助手壳」。",[11,1173,1174],{},"我不是要默认「关盖后云电脑继续干」的云队友。本机现场执行就意味着电脑合上，活也停；换来的是文件、工具和隐私边界更清楚。",[11,1176,1177],{},"我也不是要做一个技能全家桶替代多实体分工。那会重新引入上下文污染。窄任务就该找窄 Agent。",[11,1179,1180],{},"第一阶段也不急着做计费、审核、手机端那一整套增长件。先把「分人办事 + 能组队 + 本机能干活」这条体验走通。",[33,1182,1183],{"id":1183},"先把关系模型立住",[11,1185,1186],{},"这是 AgentUp 的第一篇创造日志。今天想先记清楚的，不是功能清单，而是为什么动手。",[11,1188,1189],{},"私人助理越装越臃肿，这件事我已经受够了。Agents 正在变专、变深，也开始像团队合作。我希望用 AI 的方式，能更接近现在找人办事的方式。",[11,1191,1120],{},{"title":156,"searchDepth":157,"depth":157,"links":1193},[1194,1195,1196,1197,1198],{"id":1037,"depth":160,"text":1038},{"id":1057,"depth":160,"text":1058},{"id":1097,"depth":160,"text":1098},{"id":1168,"depth":160,"text":1168},{"id":1183,"depth":160,"text":1183},"2026-10-22","个人助手型 Agent 容易变成全家桶。Agent 正在变专、变深，也能组队。AgentUp 想做的是：像用微信那样用 AI。",[1202,1205,1208,1211],{"q":1203,"a":1204},"AgentUp 想解决什么问题？","个人助手型 Agent 把论文、代码、日程、天气全塞进同一个实例后，上下文会搅成一团，窄任务也容易答偏。AgentUp 想把关系做成通讯录：一个助手管一类事，跨界再拉队友。",{"q":1206,"a":1207},"AgentUp 的 slogan 是什么意思？","「像用微信那样用 AI」指的是入口和关系像找人办事：加进通讯录、长期可聊、找谁办什么事上下文分开；不是再开一个开发工具窗，也不是往同一个全能助手里继续堆技能。",{"q":1209,"a":1210},"AgentUp 和常见私人助理、技能商店有什么不同？","私人助理默认给同一个实例无限装能力；技能商店常是复制一份私人模板。AgentUp 更接近招募专业实体、分人办事，需要时再协同，而不是把所有能力堆进一个人设。",{"q":1212,"a":1213},"现在去哪里看 AgentUp？","官网与桌面端入口：https:\u002F\u002Fagentup.me。源码开源：https:\u002F\u002Fgithub.com\u002FMorvanZhou\u002Fagentup。项目仍在持续进化中。",{},"\u002Fbuilds\u002Fagentup\u002F2026-10-22-why-agentup",{"title":1026,"description":1200},[1218,1219,1220,1221,1222,1223,1224],"AgentUp","像用微信那样用 AI","Agent 团队协作","私人助理臃肿","专业化 Agent","本机 AI 助手","独立开发 Agent IM","builds\u002Fagentup\u002F2026-10-22-why-agentup","个人助手型 Agent 越装越臃肿，窄任务也背全家桶。Agent 正在专业化、深度化，并适合团队合作。AgentUp 的核心主张是：像用微信那样用 AI，分人办事，跨界再组队。",[198,199,200],"qD8BH2J8Cdw2jgQmZFyuTrZjoLZ7mMqx9NGWAZsHPI8",1793028248146]