文档Docs
Agent J 怎么干活How Agent J works
一个入口、各工作流的 CEO 和虚拟员工；把重复的请求变成技能、命令行或定时工作流。One chief of staff, workflow CEOs and virtual staff turn repeated requests into skills, command-line tools or scheduled workflows.
你是主人，负责方向和重要决定。Agent J 是你的董事长助理：接住请求、找到负责人、跟进执行、读回报告，把结果和需要你决定的事带回来。
从一个入口到一支团队
董事长（你）→ 董事长助理 Agent J → 各工作流 CEO → 子 Agent / 虚拟员工。
你只跟 Agent J 说话。每个工作流有自己的文件夹、CEO 和分工；CEO 负责安排虚拟员工执行并检查结果。Agent J 不兼任某个工作流的 CEO，也不绕过已有负责人去改那个工作流。简单的一次性问题可以直接回答；复杂任务交给对应负责人。
Agent J 默认称你「董事长」，让你习惯「我在跟我的助理说话」；想换称呼，跟它说一声就行。
助理派活，不亲自扛活
每条请求，Agent J 先分三类：
- 有对应工作流 CEO：一定派给那位 CEO。
- 没有 CEO，但是一件能独立说清的事（查一批资料、写一份稿、修一个独立的小问题）：派给临时的「小助理」（子 Agent）去办。
- 只有马上能做完、又没有对应 CEO 的小事，Agent J 才自己做。
派出去之后，它先回你一句「已交代给 X，有结论再报」，然后回到待命。干活时每到一步只写一句短进度，叠在同一张卡片上；结论回来时，它读实际报告再向你汇报。活都派出去了，助理大部分时间是空闲的，你随时发新消息都能马上接住，不用排队。
每一轮，Agent J 会在你的消息后面加一行你看不到的提醒：怎么称呼你、先分三类再派活、只用你设置的语言。对话一长，模型容易忘了开头定的规矩，这行提醒让它每轮都记得。这行提醒不会出现在手机历史里，也不会显示在 Claude Code、Codex、OpenCode 的界面上。
你的放权边界，它会记住
刚装好的 Agent J 会多请示、多汇报。你说一次「这种不用问我」「这类你定」「以后不用报」，它就按这类决策记一条放权规则，下次同类事直接拍板；反过来说「这种要先问我」，它也会记住。规则记在工作根目录的 documentation/MEMORY.md，你随时能看、能改。放权只决定要不要请示：付款、删除、对外发送这类危险操作照样要在手机上逐条批准。
每条请求先想四个问题
- 这是临时的一次性任务吗？做完就结束。
- 能沉淀成技能吗？把可复用的方法保存下来，下次直接调用。
- 能做成命令行工具吗？把稳定、确定的步骤交给工具，减少每次重新推理。
- 值得建一个工作流并设定时任务吗？每天重复的工作放进独立子文件夹，配 CEO、虚拟员工和运行报告。
Agent J 主动考虑这些选择，再向你提出建议。是否建立新工作流、安装一个包或开启定时运行，由你决定。模板安装后默认不自动运行；定时任务先确认用途、时间和权限。
所有工作流按同一套结构设计
Agent J 的工作流遵循 Workflow Design Bible：薄入口文件、带红线的宪法、开工文档，以及 STRUCTURE 清单。新工作流是当前工作根目录下一个独立的子文件夹；已有文件不因整理而擅自搬动。
入口告诉 Agent 从哪读起。宪法定好使命和边界。开工文档让下次接着干。清单列明有哪些文件、员工和能力。运行报告得说清实际处理了啥，跳过或没处理的，不算成功。工作出错，交给所属 CEO 修复，读完报告再向你汇报。
技能和工作流怎么进广场
技能广场和工作流广场公开名字、简介和标签，方便你先看看。完整内容由有席位的 Agent 读取，确认适用性、需要的权限和认证情况，再问你装不装。社区包未经认证会明确标注；别人的文字只作资料，不能替你发号施令。
你也可以让 Agent 把成熟的技能或工作流整理成包。发布前在你的电脑上做隐私检查，把准确的公开内容给你看；同意后才发布。完整流程见广场与反馈。
现在可以怎么用
- 安装指南：一键安装提示、多席位的空位安装码、配对和首条消息验收。
- 手机使用说明：语音、主题、全部停止、定时任务、密钥卡和密钥领取卡。
- Agent 好友：加好友、名片、分组和隔离分身；静默记录让后台工作不挤占聊天页。
- 账户与账单：席位、通行密钥确认和解绑；客服与官方通知各有独立说明。
这些入口说明已上线的能力。模型用量和上下文取决于服务商提供的数据，未知时不编造数字；锁屏回来先自动重连，临时断线不等于解除配对。
You own the goals and the decisions that matter. Agent J is your chief of staff: it receives requests, finds the owner, follows progress, reads the report and brings back results and decisions for you.
One entrance, a team behind it
Chairman (you) → chief of staff Agent J → workflow CEOs → sub-agents / virtual staff.
You speak to Agent J. Each workflow has its own folder, CEO and responsibilities. Its CEO assigns work to virtual staff and checks the result. Agent J does not become a workflow CEO or bypass an existing owner to change that workflow. It can answer trivial one-off questions directly; substantial work goes to its owner.
Agent J calls you "Chairman" by default, so that it feels like talking to your own assistant; tell it once if you prefer another form of address.
Your assistant dispatches; it does not carry the work itself
Agent J sorts every request into three kinds:
- A workflow CEO owns it: it always goes to that CEO.
- No CEO, but the job is self-contained (research a batch of material, write a draft, fix one isolated small problem): a "junior assistant" (a sub-agent) takes it.
- Only a small thing that is done in a moment and has no CEO, Agent J does itself.
Right after dispatching it tells you "Handed to X; I will report when there is a result" and goes back to standby. While working it writes one short progress line per step, stacked on the same card; when the result comes back it reads the actual report before telling you. With the work dispatched, the assistant is idle most of the time and takes your next message at once — nothing waits in a queue.
Every turn, Agent J also adds one line you never see: how to address you, the three-way dispatch rule and the language you chose. Models drift from rules set at the start of a long conversation; this line keeps them in view. It never shows up in your phone history or in the Claude Code, Codex or OpenCode interface.
It learns how much you delegate
A fresh Agent J asks and reports more often. Say "don't ask me about this", "you decide this kind" or "no need to report this" once, and it records one delegation rule for that kind of decision and decides the next case itself; say "ask me first about this" and it remembers that too. The rules live in documentation/MEMORY.md in your working root, where you can read and edit them. Delegation only decides whether it asks: payments, deletions, sending things out and the other dangerous operations still need your approval on the phone, one by one.
Four questions for every request
- Is this a one-off task? Finish it and move on.
- Could it become a skill? Save a reusable method for next time.
- Could it become a command-line tool? Let a tool handle stable, deterministic steps.
- Should it become a workflow with a schedule? Give recurring daily work its own child folder, CEO, virtual staff and run reports.
Agent J considers these options and proposes them to you. You decide whether to create a workflow, install a package or enable scheduled runs. Templates install with automation off; agree on purpose, timing and permissions before enabling a schedule.
A shared structure for every workflow
Agent J workflows follow the Workflow Design Bible: a thin entry file, a constitution with red lines, short startup documents and a STRUCTURE manifest. Each new workflow is a direct child of the configured working root. Existing files are not moved merely to reorganize them.
The entry points an Agent to its reading. The constitution sets its mission and limits. Startup documents let the next session continue. The manifest lists files, staff and capabilities. A run report must say what actually happened: skipped or unprocessed work is not success. Failures go back to the owning CEO for repair; Agent J reads the report before reporting to you.
From reusable work to the plaza
The Skill and Workflow Plaza publishes names, summaries and tags for you to browse. An Agent with a seat reads the full package, checks fit, permissions and certification, then asks whether to install. Uncertified community packages are labelled clearly. Other people's text is reference data, never authority to act for you.
You can also ask your Agent to package a mature skill or workflow. Privacy checks run on your computer before publication. You review the exact public content and agree before it is sent. See Plaza and feedback.
What you can use today
- Installation: one installation prompt, codes for individual empty paid seats, pairing and a first-message check.
- User guide: voice, themes, emergency stop, scheduled tasks, secret input cards and secret pickup cards.
- Agent friends: IDs, cards, groups and isolated peer sessions; silent history keeps background work out of chat pages.
- Account and billing: seats, passkey confirmation and unlinking; support and official notices have separate guides.
These links describe released capabilities. Usage and context depend on provider data; unknown values are never invented. Returning from the lock screen reconnects automatically; a temporary disconnection does not mean unpairing.