有些事不是一次性的：每周看一遍竞品、每天整理一份资料、每月对一次账。在 Agent J 里，你只要用一句话说出来，比如「做每周竞品简报」。剩下的由你的 Agent 来办：先查它现在会什么、还缺什么，把真正需要你决定的事集中到一张开工卡上；你确认一次，它就按时自己跑、自己交付。

这里没有新按钮。所有事都从聊天里一句话开始。

一次完整的过程

1. 你说一句话。 例如「每周一早上 9 点，帮我做一份三家竞品的公开资料简报」。
2. Agent 自查。 它会看这台电脑上已经装好的能力：能不能读公开网页、能不能把文件存到电脑上、你的手机有没有配好，还有已经有哪些工作流。能自己检查的，它都先查一遍，不来问你。
3. 一张开工卡。 真正要你决定的事（比如关注哪三家、谁来接收）最多五项，放在同一张卡上。卡上还写着这次要做什么、交付到哪里、什么时候自动运行。
4. 你点「确认并开始」。 如果是新工作流，或者要每周自动运行，这一下同时确认任务书并启用定时。
5. 以后按时自己跑。 每次运行完，负责这个工作流的 CEO 会写一份报告。Agent 读过报告、核对过产物，才会告诉你已经完成。

开工卡上有什么

• 这次要做什么：目标、交付物（比如 reports/weekly.md）、多久运行一次。
• 需要你选：最多五项。可以是选项，也可以写一句话。选过的内容以后每次沿用。
• 还要配置的：例如还没登录的网站、还没存的服务。这些不在卡上填：要密钥时，手机会另外弹出专用的密钥卡，Agent 看不到密钥本身。
• 可选项：没有也能完成，比如付费调研服务。缺了也不会挡住任务，Agent 会改用免费的公开资料。
• 点「确认并开始」等于同意什么：按这份任务书新建工作流、按时运行、沿用你的选择。

确认不等于同意付款、删除、对外发送、改密钥或改价。这五类事以后每次都会单独在手机上问你。把结果推到你自己已配对的手机上，不算「对外发送」。

点「返回修改」就不会开工。直接在聊天里说想改什么就行。卡片大约 30 分钟内有效，过期后什么都不会发生。

「你会什么」

在聊天里问一句「你会什么」，手机上会弹出能力页：哪些已就绪、哪些需要配置，可以筛选，点一项能看说明。能力会定期自动复验，比如网页读取每 15 分钟、工具每 7 天。这只是检查，不会因此再来问你。

清单还会列出专用浏览器、已配置的网站和 Google 工具。网站登录检查超过 15 分钟就过期，浏览器停止时相关能力不可用。装好 Google 工具不等于完成接口授权；接口授权的一键引导在 0.18.1。

什么时候它还会来问你

• 五类高风险操作：付款、删除、对外发送、改密钥、改价，每次都单独问。
• 真正缺少你的决定或授权，比如某个网站要你本人登录。这时只停下依赖它的那部分，别的照常继续，事后把缺口集中告诉你。
• 你改了目标、交付或定时。小的措辞调整不会重新问你。

需要知道的限制

• 电脑要开着、联网。 电脑休眠时任务不会跑，Agent 不会假装云端还在继续。参见让电脑保持唤醒。
• 连接器（MCP）只列出，不会自动试运行：启动第三方程序去测试本身就是一种操作，所以它们显示为「待检查」。
• 旧版手机页面看不到开工卡，只会收到一行提示，刷新页面即可。

主机重启后

主机服务重启后，未完成的工作流显示中断状态，已有报告保留，不会自动重放或复用旧审批。若无法确认旧执行是否结束，先核对报告和发送、发布等外部操作，再由主人重新启用该工作流；手动重跑会得到新的运行编号。停机期间错过的定时任务仍不补跑。

隐私

能力清单、任务书、你的选择、报告都只存在这台电脑上，不会进账号后台，也不会进运营方的数据库。手机和电脑之间全程端到端加密，中继服务器只能看到密文。

给你的 Agent 看（命令参考）

这些命令由主 Agent 通过内置的 agentj-capability 技能调用，你不需要自己敲：

agentj capability list|sync|check [--json]      能力清单 / 发现并检查 / 复验
agentj capability show                           把「我的 Agent 会什么」发到主人手机
agentj capability digest --dir <工作流>           任务书摘要与问题检查
agentj capability prepare --workflow <id> [--asks <文件>]   开工卡（≤5 项）
agentj capability result <卡片 id> [--wait]
agentj capability dispatch --workflow <id>       派给工作流 CEO（已在跑就不重复）
agentj capability report-check --dir <工作流> --verdict <ok|attention|fail>

任务书是工作流目录里的 brief.json（字段固定），CEO 的报告是 reports/report.json。没有报告、验收没通过或产物读不到，都不算完成。

Long tasks: say it once, it runs on its own

Some work repeats: a weekly look at competitors, a daily digest, a monthly reconciliation. With Agent J you say it once in chat — for example "set up a weekly competitor brief". Your Agent checks what it can already do, asks everything that is genuinely yours to decide on one start card, and after a single confirmation it runs and delivers on schedule.

There is no new button. Everything starts with a sentence in chat.

How it goes

1. You say it. "Every Monday at 9:00, a brief on three competitors from public sources."
2. Your Agent checks itself. Can it read public web pages, save files on this computer, reach your paired phone, which workflows exist already? Whatever it can check, it checks — without asking you.
3. One start card. At most five things that only you can decide (which three competitors, who receives it), on one card, together with what will be done, where it is delivered and when it runs.
4. You tap Confirm and start. For a new workflow or a weekly schedule, this one tap confirms the brief and turns the schedule on.
5. It runs on its own. After every run the workflow's CEO writes a report. Your Agent reads the report and checks the results before it tells you "done".

What the start card shows

• What this does: the goal, the deliverable (for example reports/weekly.md) and how often it runs.
• Your choices: at most five — options to tap or a short answer. They are reused every time.
• Still to set up: a site you have not signed in to, a service not saved yet. Nothing to type on the card: a key opens the dedicated secret card, and your Agent never sees the value.
• Optional: things the task works without, such as a paid research service. Missing ones never block — your Agent uses free public sources instead.
• What Confirm and start means: create the workflow from this brief, run it on schedule, reuse your choices.

Confirming does not approve paying, deleting, sending anything outside, changing keys or changing prices: those still ask you on the phone, one by one. Sending the result to your own paired phone does not count as sending anything outside.

Change something sends the card back and nothing starts — just say in chat what to change. A card is valid for about 30 minutes; when it expires nothing happens.

"What can you do"

Ask "what can you do" in chat and a capability page opens on your phone: what is ready, what needs setup, filterable, tap an item for details. Capabilities are re-checked automatically (web reading every 15 minutes, tools every 7 days) — a re-check never becomes a new question for you.

The list also includes the dedicated browser and its configured sites, plus the Google tool. Site sign-ins expire after 15 minutes and a stopped browser is unavailable. Installing the Google tool does not authorize Google APIs; guided API authorization arrives in 0.18.1.

When it still asks you

• The five high-risk actions: paying, deleting, sending outside, key changes and price changes — every time, separately.
• A decision or authorization that is genuinely missing, such as a site only you can sign in to. Only the part that depends on it waits; everything else continues, and the gap is reported in one go.
• You changed the goal, deliverable or schedule. A small wording change does not ask again.

Limits worth knowing

• The computer must be on and online. A sleeping computer runs nothing, and your Agent will not pretend the cloud carries on. See Keep the computer awake.
• Connectors (MCP) are listed but never test-started: starting a third-party program to test it is itself an action, so they show as "To check".
• An older phone page cannot show the start card; it gets one line asking you to reload the page.

After a host restart

After a host restart, unfinished workflows show as interrupted and existing reports remain. Old runs and approvals are never replayed. If the old execution may still be active, check its reports and external effects, then have the owner enable the workflow again; a manual rerun gets a new run ID. Missed schedules still have no catch-up.

Privacy

The capability list, briefs, your choices and reports stay on this computer only — not in the account dashboard, not in the operator's database. Phone and computer talk end to end encrypted; the relay sees ciphertext only.

For your Agent (command reference)

The main Agent runs these through its built-in agentj-capability skill; you never type them:

agentj capability list|sync|check [--json]      inventory / discover and check / re-check
agentj capability show                           put "What my Agent can do" on the owner's phone
agentj capability digest --dir <workflow>        brief digests and problems
agentj capability prepare --workflow <id> [--asks <file>]   the start card (≤ 5 items)
agentj capability result <card id> [--wait]
agentj capability dispatch --workflow <id>       hand it to the workflow CEO (never twice)
agentj capability report-check --dir <workflow> --verdict <ok|attention|fail>

The brief is brief.json in the workflow folder (fixed fields); the CEO's report is reports/report.json. No report, unverified acceptance or an unreadable result is not done.