派出去 · Cloud Agents、CLI 与自动化Send It Out · Cloud Agents, CLI, and Automations.
有些活不用你守着：交给云端 agent，它在自己的机器上跑，你去忙别的；放进 CLI，它在终端和 CI 里无人值守地跑；设条自动化，它按时间或事件自己触发。Cursor 把同一个 agent 铺到了网页、iOS、Slack、GitHub 上，还给了一套 SDK 让你编排它。这一章讲哪类活值得派出去（包括 Bugbot 那种自动 PR 评审），哪类还是留在你眼前。Some work doesn't need you watching: hand it to a cloud agent and it runs on its own machine while you do something else; drop it into the CLI and it runs unattended in the terminal and CI; set an automation and it fires on a schedule or an event. Cursor spreads the same agent across web, iOS, Slack, and GitHub — and ships an SDK to orchestrate it. Which work is worth sending out — including Bugbot's automatic PR review — and which should stay in front of you.
有些活不用你守着：交给云端 agent，它在自己的机器上跑，你去忙别的；放进 CLI，它在终端和 CI 里无人值守地跑；设条自动化，它按时间或事件自己触发。Some work doesn't need you watching: hand it to a cloud agent and it runs on its own machine while you do something else; drop it into the CLI and it runs unattended in the terminal and CI; set an automation and it fires on a schedule or an event.Cursor 把同一个 agent 铺到了网页、手机、Slack、GitHub 上。前面几章的 agent 是「陪你做」的，这一章讲「派出去做」的那一面 —— 哪类活值得交出去（包括 Bugbot 那种自动 PR 评审），哪类还是留在你眼前。Cursor spreads the same agent across web, mobile, Slack, and GitHub. The agent in earlier chapters works with you; this chapter is the other face — sending it out — which work is worth handing off (including Bugbot's automatic PR review), and which should stay in front of you.
— I
Cloud Agents：派出去，它在自己机器上跑Cloud Agents: Hand It Off, It Runs on Its Own Machine.
本地 agent 占着你的机器和注意力；云端 agent 不占。A local agent ties up your machine and your attention; a cloud agent doesn't.云端 agent（原 Background Agents）各自有一台隔离 VM，可以并行跑好几个，还带 computer use 沙箱去验证自己的产出。1注 1Note 1Cursor Docs · Cloud Agents —— 原 Background Agents；每个 agent 一台隔离 VM、装完整开发环境，可随便并行，能在自己的桌面 / 浏览器里验证产出；发起入口：iOS app、cursor.com/agents 网页（Android 装 PWA）、桌面、Slack / Linear @cursor、GitHub / Bitbucket 评论 @cursor、API；本地会话里 /in-cloud 把下一个任务丢给云端 subagent，/babysit 让它盯 PR 到能合并。云端一律跑 Max Mode。截至 2026-07-10。Cursor Docs · Cloud Agents — formerly Background Agents; each gets an isolated VM with a full dev environment, runs in parallel at will, and verifies output on its own desktop / browser; start from the iOS app, cursor.com/agents on the web (PWA on Android), desktop, @cursor in Slack / Linear, @cursor comments on GitHub / Bitbucket, or the API; in a local session, /in-cloud hands the next task to a cloud subagent and /babysit has one shepherd a PR to merge. Cloud runs always use Max Mode. As of 2026-07-10.发起入口铺得很开：网页、iOS app、桌面、Slack / Linear 里 @cursor、GitHub / Bitbucket 的 PR 评论里 @cursor，也能把本地或 CLI 上跑着的活 handoff 到云端继续。它克隆你的仓库、在自己机器上改完、交一个 PR 回来。Cloud agents (formerly Background Agents) each get an isolated VM, run several in parallel, and carry a computer-use sandbox to validate their own output.1注 1Note 1Cursor Docs · Cloud Agents —— 原 Background Agents；每个 agent 一台隔离 VM、装完整开发环境，可随便并行，能在自己的桌面 / 浏览器里验证产出；发起入口：iOS app、cursor.com/agents 网页（Android 装 PWA）、桌面、Slack / Linear @cursor、GitHub / Bitbucket 评论 @cursor、API；本地会话里 /in-cloud 把下一个任务丢给云端 subagent，/babysit 让它盯 PR 到能合并。云端一律跑 Max Mode。截至 2026-07-10。Cursor Docs · Cloud Agents — formerly Background Agents; each gets an isolated VM with a full dev environment, runs in parallel at will, and verifies output on its own desktop / browser; start from the iOS app, cursor.com/agents on the web (PWA on Android), desktop, @cursor in Slack / Linear, @cursor comments on GitHub / Bitbucket, or the API; in a local session, /in-cloud hands the next task to a cloud subagent and /babysit has one shepherd a PR to merge. Cloud runs always use Max Mode. As of 2026-07-10. The entrances are everywhere: web, the iOS app, desktop, @cursor in Slack / Linear, @cursor in a GitHub / Bitbucket PR comment — and you can hand off work running locally or in the CLI to continue in the cloud. It clones your repo, finishes on its own machine, and hands back a PR.「派出去」也不非得整单外包：3.7 起，本地会话里 /in-cloud 就能把下一个任务丢给一个云端 subagent，它在自己的 VM 和分支上跑，你的本地工作区不被占用；/babysit 则让一个云端 agent 盯住某个 PR，回评论、解冲突、修挂掉的检查，直到它能合并。1注 1Note 1Cursor Docs · Cloud Agents —— 原 Background Agents；每个 agent 一台隔离 VM、装完整开发环境，可随便并行，能在自己的桌面 / 浏览器里验证产出；发起入口：iOS app、cursor.com/agents 网页（Android 装 PWA）、桌面、Slack / Linear @cursor、GitHub / Bitbucket 评论 @cursor、API；本地会话里 /in-cloud 把下一个任务丢给云端 subagent，/babysit 让它盯 PR 到能合并。云端一律跑 Max Mode。截至 2026-07-10。Cursor Docs · Cloud Agents — formerly Background Agents; each gets an isolated VM with a full dev environment, runs in parallel at will, and verifies output on its own desktop / browser; start from the iOS app, cursor.com/agents on the web (PWA on Android), desktop, @cursor in Slack / Linear, @cursor comments on GitHub / Bitbucket, or the API; in a local session, /in-cloud hands the next task to a cloud subagent and /babysit has one shepherd a PR to merge. Cloud runs always use Max Mode. As of 2026-07-10.手机那头也接上了：iOS app（beta，官方公告称 public beta，各付费档可用）能从手机发起、跟进、直接合并 agent 的 PR；Remote Control 让你人在外面、继续指挥跑在你电脑上的那个 agent；Live Activities 把进度放在锁屏上。5注 5Note 5Cursor Docs · Cursor for iOS —— iOS app 处于 beta（changelog 称 public beta，2026-06-29 起），各付费档可用：从手机发起、跟进、合并云端 agent 的 PR，语音描述任务；Remote Control 从手机继续指挥跑在你电脑上的 agent（Teams / Enterprise 需管理员开启）；Live Activities 锁屏看进度；iPhone（iOS 26+）、英文界面，Android 计划中。截至 2026-07-10。Cursor Docs · Cursor for iOS — the iOS app is in beta (the changelog calls it a public beta, since 2026-06-29), on all paid plans: start, follow, and merge cloud agents' PRs from your phone, with voice input; Remote Control keeps directing an agent running on your computer from your phone (admin-enabled on Teams / Enterprise); Live Activities show progress on the lock screen; iPhone (iOS 26+), English UI, Android planned. As of 2026-07-10.判断不变：手机适合查进度、批结果、给下一步指令，不适合当主工作台。'Sending out' needn't be all-or-nothing either: since 3.7, /in-cloud in a local session hands the next task to a cloud subagent that runs on its own VM and branch, leaving your local workspace free; /babysit has a cloud agent shepherd a PR — answering comments, resolving conflicts, fixing failed checks — until it can merge.1注 1Note 1Cursor Docs · Cloud Agents —— 原 Background Agents；每个 agent 一台隔离 VM、装完整开发环境，可随便并行，能在自己的桌面 / 浏览器里验证产出；发起入口：iOS app、cursor.com/agents 网页（Android 装 PWA）、桌面、Slack / Linear @cursor、GitHub / Bitbucket 评论 @cursor、API；本地会话里 /in-cloud 把下一个任务丢给云端 subagent，/babysit 让它盯 PR 到能合并。云端一律跑 Max Mode。截至 2026-07-10。Cursor Docs · Cloud Agents — formerly Background Agents; each gets an isolated VM with a full dev environment, runs in parallel at will, and verifies output on its own desktop / browser; start from the iOS app, cursor.com/agents on the web (PWA on Android), desktop, @cursor in Slack / Linear, @cursor comments on GitHub / Bitbucket, or the API; in a local session, /in-cloud hands the next task to a cloud subagent and /babysit has one shepherd a PR to merge. Cloud runs always use Max Mode. As of 2026-07-10. The phone is wired in too: the iOS app (beta; the announcement calls it a public beta, on all paid plans) starts, follows, and merges agents' PRs from your phone; Remote Control keeps directing the agent running on your computer while you're out; Live Activities put progress on the lock screen.5注 5Note 5Cursor Docs · Cursor for iOS —— iOS app 处于 beta（changelog 称 public beta，2026-06-29 起），各付费档可用：从手机发起、跟进、合并云端 agent 的 PR，语音描述任务；Remote Control 从手机继续指挥跑在你电脑上的 agent（Teams / Enterprise 需管理员开启）；Live Activities 锁屏看进度；iPhone（iOS 26+）、英文界面，Android 计划中。截至 2026-07-10。Cursor Docs · Cursor for iOS — the iOS app is in beta (the changelog calls it a public beta, since 2026-06-29), on all paid plans: start, follow, and merge cloud agents' PRs from your phone, with voice input; Remote Control keeps directing an agent running on your computer from your phone (admin-enabled on Teams / Enterprise); Live Activities show progress on the lock screen; iPhone (iOS 26+), English UI, Android planned. As of 2026-07-10. The judgment stands: a phone is for checking progress, approving results, and giving the next instruction — not for being the main workbench.判断：陪你做的活留在眼前，做完通知的活派出去。它适合可并行、定义清楚、跑错能回滚的任务；不适合探索性的、你自己都没想清楚的改动 —— 那种你不在场，它容易产出一堆你没有心智模型、又得硬着头皮审的代码。它的 computer use 只在自己沙箱里（验证它造的东西），不碰你本机，这条线第 10 章再收。The call: work-with-me stays in front of you, done-tell-me gets sent out. It suits parallelizable, well-defined, rollback-able tasks; not exploratory changes you haven't thought through — with you absent, those tend to produce a pile of code you have no mental model for and must review anyway. Its computer use stays in its own sandbox (validating what it built), never your machine — that line closes in Chapter 10.
提示词PromptCloud Agent
Cloud Agent 任务：
目标：<做成什么>
仓库/分支：<repo + base branch>
范围：只改 <paths>
环境：运行 <setup/test commands>
验收：
- <测试命令必须通过>
- <PR 描述必须列出风险和回滚>
禁止：生产操作、读取 .env*、改权限、无测试大改。
完成后开 PR，并附日志 / 截图 / artifact。Cloud Agent task:
Goal: <what to finish>
Repo/branch: <repo + base branch>
Scope: modify only <paths>
Environment: run <setup/test commands>
Acceptance:
- <test command must pass>
- <PR description must list risks and rollback>
Forbidden: production actions, reading .env*, permission changes, large untested rewrites.
Open a PR with logs / screenshots / artifacts.
Cloud Agent 任务：
目标：把所有日期格式从 MM/DD/YYYY 统一改成 ISO 8601（YYYY-MM-DD）
仓库/分支：acme/billing-service，base 分支 main
范围：只改 src/utils/date.ts 和 src/components/**
环境：运行 npm ci && npm run test:date
验收：
- npm run test:date 必须全绿
- PR 描述必须列出风险和回滚（指出哪些展示位会变）
禁止：生产操作、读取 .env*、改权限、无测试大改。
完成后开 PR，并附日志 / 截图 / artifact。Cloud Agent task:
Goal: standardize all date formats from MM/DD/YYYY to ISO 8601 (YYYY-MM-DD)
Repo/branch: acme/billing-service, base branch main
Scope: modify only src/utils/date.ts and src/components/**
Environment: run npm ci && npm run test:date
Acceptance:
- npm run test:date must be all green
- PR description must list risks and rollback (note which display spots change)
Forbidden: production actions, reading .env*, permission changes, large untested rewrites.
Open a PR with logs / screenshots / artifacts.
选择Choice
适合Good for
不适合Bad for
本地 AgentLocal Agent
你要边看边改、随时打断、审每一步You want to watch, interrupt, and review each step
明确、可回滚、跑完给 PR 的任务Clear, rollbackable tasks that should return a PR
敏感账号操作、无测试大改、探索性设计Sensitive account actions, large untested changes, exploratory design
Cursor CLI
脚本、CI、headless review、非 IDE 场景Scripts, CI, headless review, non-IDE contexts
需要大量交互式 UI 判断的工作Work needing heavy interactive UI judgment
Cursor SDK
把 agent 编成你产品或流水线的部件：CI 修复机器人、批量任务、自定义工具Agents as parts of your product or pipeline: CI fix bots, batch jobs, custom tools
一次性的活、CLI 一行就能解决的活One-off work, or anything one CLI line already solves
— II
CLI 与自动化：无人值守地跑CLI and Automations: Run Unattended.
能进脚本的 agent，才能在你睡觉时干活。An agent that fits a script is one that works while you sleep.agent 把 agent 带进终端：-p 非交互、--output-format json 给结构化输出，可以塞进脚本和 CI / GitHub Actions；消息前加 &，任务就推给云端接着跑。2注 2Note 2Cursor Docs · CLI —— agent 把 agent 带到终端；-p 非交互，--output-format json / stream-json 结构化输出，可进脚本与 CI / GitHub Actions；支持 Plan / Ask 模式；消息前缀 & 即把任务推给 Cloud Agent 继续跑。截至 2026-07-10。Cursor Docs · CLI — agent brings the agent to the terminal; -p for non-interactive runs, --output-format json / stream-json for structured output, fits scripts and CI / GitHub Actions; supports Plan / Ask modes; prefix a message with & to push the task to a Cloud Agent. As of 2026-07-10.Automations 则让云端 agent 按计划或事件自己触发 —— 事件面很宽：GitHub / GitLab / Bitbucket 的 PR 与评论、CI 跑完、Slack 消息（甚至指定一个表情当扳机）、webhooks、Linear。3注 3Note 3Cursor Docs · Automations —— 让 Cloud Agents 按计划或事件运行，事件源含 GitHub、GitLab、Bitbucket、Slack（含表情触发）、webhooks、Linear 等；本地会话里用 /automate 一句话建自动化，Cursor 替你配触发器、指令与工具；computer use 对每条自动化默认开启，可要求它交 demo / artifact；自动化一律以 Max Mode 计费。截至 2026-07-10。Cursor Docs · Automations — run Cloud Agents on a schedule or on events from GitHub, GitLab, Bitbucket, Slack (including an emoji trigger), webhooks, Linear, and more; /automate in a local session builds one from a plain-language description, with Cursor configuring triggers, instructions, and tools; computer use is on by default for every automation, so you can ask for a demo / artifact; automations always bill at Max Mode. As of 2026-07-10.建一条也不再需要你会配：3.8 起在本地会话里 /automate，一句话描述要自动化什么，它替你配好触发器、指令和工具；自动化拉起的云端 agent 默认带 computer use，你可以直接在指令里要求它交一段 demo 录屏当作业。3注 3Note 3Cursor Docs · Automations —— 让 Cloud Agents 按计划或事件运行，事件源含 GitHub、GitLab、Bitbucket、Slack（含表情触发）、webhooks、Linear 等；本地会话里用 /automate 一句话建自动化，Cursor 替你配触发器、指令与工具；computer use 对每条自动化默认开启，可要求它交 demo / artifact；自动化一律以 Max Mode 计费。截至 2026-07-10。Cursor Docs · Automations — run Cloud Agents on a schedule or on events from GitHub, GitLab, Bitbucket, Slack (including an emoji trigger), webhooks, Linear, and more; /automate in a local session builds one from a plain-language description, with Cursor configuring triggers, instructions, and tools; computer use is on by default for every automation, so you can ask for a demo / artifact; automations always bill at Max Mode. As of 2026-07-10.装上很简单，跑一个非交互任务也就一行：agent brings the agent into the terminal: -p for non-interactive, --output-format json for structured output, droppable into scripts and CI / GitHub Actions; prefix a message with & and the task moves to the cloud to keep running.2注 2Note 2Cursor Docs · CLI —— agent 把 agent 带到终端；-p 非交互，--output-format json / stream-json 结构化输出，可进脚本与 CI / GitHub Actions；支持 Plan / Ask 模式；消息前缀 & 即把任务推给 Cloud Agent 继续跑。截至 2026-07-10。Cursor Docs · CLI — agent brings the agent to the terminal; -p for non-interactive runs, --output-format json / stream-json for structured output, fits scripts and CI / GitHub Actions; supports Plan / Ask modes; prefix a message with & to push the task to a Cloud Agent. As of 2026-07-10. Automations fire a cloud agent on a schedule or an event — and the event surface is wide: PRs and comments on GitHub / GitLab / Bitbucket, a CI run finishing, Slack messages (even a designated emoji as the trigger), webhooks, Linear.3注 3Note 3Cursor Docs · Automations —— 让 Cloud Agents 按计划或事件运行，事件源含 GitHub、GitLab、Bitbucket、Slack（含表情触发）、webhooks、Linear 等；本地会话里用 /automate 一句话建自动化，Cursor 替你配触发器、指令与工具；computer use 对每条自动化默认开启，可要求它交 demo / artifact；自动化一律以 Max Mode 计费。截至 2026-07-10。Cursor Docs · Automations — run Cloud Agents on a schedule or on events from GitHub, GitLab, Bitbucket, Slack (including an emoji trigger), webhooks, Linear, and more; /automate in a local session builds one from a plain-language description, with Cursor configuring triggers, instructions, and tools; computer use is on by default for every automation, so you can ask for a demo / artifact; automations always bill at Max Mode. As of 2026-07-10. Building one no longer requires knowing the config either: since 3.8, /automate in a local session takes a plain-language description and sets up the triggers, instructions, and tools for you; the cloud agents an automation launches carry computer use by default, so your instructions can demand a screen-recording demo as proof of work.3注 3Note 3Cursor Docs · Automations —— 让 Cloud Agents 按计划或事件运行，事件源含 GitHub、GitLab、Bitbucket、Slack（含表情触发）、webhooks、Linear 等；本地会话里用 /automate 一句话建自动化，Cursor 替你配触发器、指令与工具；computer use 对每条自动化默认开启，可要求它交 demo / artifact；自动化一律以 Max Mode 计费。截至 2026-07-10。Cursor Docs · Automations — run Cloud Agents on a schedule or on events from GitHub, GitLab, Bitbucket, Slack (including an emoji trigger), webhooks, Linear, and more; /automate in a local session builds one from a plain-language description, with Cursor configuring triggers, instructions, and tools; computer use is on by default for every automation, so you can ask for a demo / artifact; automations always bill at Max Mode. As of 2026-07-10. Installing is simple, and a non-interactive task is one line:
装上 CLI，跑一个非交互任务（输出结构化 JSON）Install the CLI, run one non-interactive task (structured JSON output)
提示词PromptCursor CLI
agent -p "<任务>"
--model <model-or-auto>
--mode agent
--output-format json
约束：
- 只改 <paths>
- 不提交、不 push、不发 PR
- 所有写入先生成 diff
- 输出 JSON 包含 changedFiles、commandsRun、tests、risks。agent -p "<task>"
--model <model-or-auto>
--mode agent
--output-format json
Constraints:
- Modify only <paths>
- Do not commit, push, or open PRs
- Generate a diff before writes
- Output JSON with changedFiles, commandsRun, tests, risks.
agent -p "把 src/api 下所有 console.log 换成 logger.debug，并在文件顶部补上 logger 的 import"
--model auto
--mode agent
--output-format json
约束：
- 只改 src/api/**
- 不提交、不 push、不发 PR
- 所有写入先生成 diff
- 输出 JSON 包含 changedFiles、commandsRun、tests、risks。agent -p "replace every console.log under src/api with logger.debug and add the logger import at the top of each file"
--model auto
--mode agent
--output-format json
Constraints:
- Modify only src/api/**
- Do not commit, push, or open PRs
- Generate a diff before writes
- Output JSON with changedFiles, commandsRun, tests, risks.
CI 里用 `CURSOR_API_KEY`，不要把 token 写进 workflow。Use `CURSOR_API_KEY` in CI; never hardcode the token in workflow files.
用权限配置限制读写路径和 shell 命令。Use permission config to restrict read/write paths and shell commands.
失败要有日志、artifact、通知；成功也要有人 review。Failures need logs, artifacts, and notification; success still needs human review.
边界：无人值守适合可验证、低风险、跑错能回滚的活 —— 定时整理数据、跑一份报告、批量处理文件。改动大、要判断、出错代价高的，别全自动；自动化的前提是「做完算什么样」你已经定死，机器只是去够这个标准。The boundary: unattended suits verifiable, low-risk, rollback-able work — tidy data on a schedule, run a report, batch-process files. Big, judgment-heavy, high-cost-of-error changes, don't fully automate; automation presumes you've already fixed 'what done looks like,' and the machine only reaches for that bar.
无人值守禁止：没有测试命令、没有日志、没有回滚办法的代码大改。Unattended forbidden: large code changes with no test command, logs, or rollback path.
— III
Bugbot：把机械评审那半交出去Bugbot: Hand Off the Mechanical Half of Review.
PR review 里最累的是机械检查，这部分该交出去。4注 4Note 4Cursor Docs · Bugbot —— 在 GitHub / GitLab / Bitbucket 的 PR 上自动评审 diff，留下解释与修复建议；会读已有评论避免重复；效果档 Default / High / Custom；usage-based 计费；可用 cursor review / bugbot run 手动触发，推送前可在本地 /review 先跑。官方 changelog（2026-06-10）称 Bugbot 现由 Composer 2.5 驱动、平均评审约 90 秒。截至 2026-07-10。Cursor Docs · Bugbot — automatically reviews PR diffs on GitHub / GitLab / Bitbucket, leaving explanations and fix suggestions; reads existing comments to avoid duplicates; Default / High / Custom effort levels; usage-based billing; trigger manually with cursor review / bugbot run, or run /review locally before pushing. Per the official changelog (2026-06-10), Bugbot is now powered by Composer 2.5, averaging ~90-second reviews. As of 2026-07-10.The tiring part of PR review is the mechanical pass — that's the part to hand off.4注 4Note 4Cursor Docs · Bugbot —— 在 GitHub / GitLab / Bitbucket 的 PR 上自动评审 diff，留下解释与修复建议；会读已有评论避免重复；效果档 Default / High / Custom；usage-based 计费；可用 cursor review / bugbot run 手动触发，推送前可在本地 /review 先跑。官方 changelog（2026-06-10）称 Bugbot 现由 Composer 2.5 驱动、平均评审约 90 秒。截至 2026-07-10。Cursor Docs · Bugbot — automatically reviews PR diffs on GitHub / GitLab / Bitbucket, leaving explanations and fix suggestions; reads existing comments to avoid duplicates; Default / High / Custom effort levels; usage-based billing; trigger manually with cursor review / bugbot run, or run /review locally before pushing. Per the official changelog (2026-06-10), Bugbot is now powered by Composer 2.5, averaging ~90-second reviews. As of 2026-07-10.Bugbot 在 GitHub / GitLab / Bitbucket 的 PR 上自动评审 diff，留下解释和修复建议，会读已有评论避免重复；效果分 Default / High / Custom 三档，usage-based 计费，也能用 cursor review / bugbot run 手动触发。4注 4Note 4Cursor Docs · Bugbot —— 在 GitHub / GitLab / Bitbucket 的 PR 上自动评审 diff，留下解释与修复建议；会读已有评论避免重复；效果档 Default / High / Custom；usage-based 计费；可用 cursor review / bugbot run 手动触发，推送前可在本地 /review 先跑。官方 changelog（2026-06-10）称 Bugbot 现由 Composer 2.5 驱动、平均评审约 90 秒。截至 2026-07-10。Cursor Docs · Bugbot — automatically reviews PR diffs on GitHub / GitLab / Bitbucket, leaving explanations and fix suggestions; reads existing comments to avoid duplicates; Default / High / Custom effort levels; usage-based billing; trigger manually with cursor review / bugbot run, or run /review locally before pushing. Per the official changelog (2026-06-10), Bugbot is now powered by Composer 2.5, averaging ~90-second reviews. As of 2026-07-10.它现在由 Cursor 自家的 Composer 2.5 驱动，官方称平均一次评审约 90 秒；推送前还能在本地 /review 先跑一遍 —— 开 PR 时 Bugbot 认得这份已审过的 diff，不再重复评审。4注 4Note 4Cursor Docs · Bugbot —— 在 GitHub / GitLab / Bitbucket 的 PR 上自动评审 diff，留下解释与修复建议；会读已有评论避免重复；效果档 Default / High / Custom；usage-based 计费；可用 cursor review / bugbot run 手动触发，推送前可在本地 /review 先跑。官方 changelog（2026-06-10）称 Bugbot 现由 Composer 2.5 驱动、平均评审约 90 秒。截至 2026-07-10。Cursor Docs · Bugbot — automatically reviews PR diffs on GitHub / GitLab / Bitbucket, leaving explanations and fix suggestions; reads existing comments to avoid duplicates; Default / High / Custom effort levels; usage-based billing; trigger manually with cursor review / bugbot run, or run /review locally before pushing. Per the official changelog (2026-06-10), Bugbot is now powered by Composer 2.5, averaging ~90-second reviews. As of 2026-07-10.它是「派出去」的一种形态：你不必盯着每个 PR 做那一遍机械扫描。Bugbot auto-reviews PR diffs on GitHub / GitLab / Bitbucket, leaves explanations and fix suggestions, and reads existing comments to avoid duplicates; effort comes in Default / High / Custom, billing is usage-based, and you can trigger it manually with cursor review / bugbot run.4注 4Note 4Cursor Docs · Bugbot —— 在 GitHub / GitLab / Bitbucket 的 PR 上自动评审 diff，留下解释与修复建议；会读已有评论避免重复；效果档 Default / High / Custom；usage-based 计费；可用 cursor review / bugbot run 手动触发，推送前可在本地 /review 先跑。官方 changelog（2026-06-10）称 Bugbot 现由 Composer 2.5 驱动、平均评审约 90 秒。截至 2026-07-10。Cursor Docs · Bugbot — automatically reviews PR diffs on GitHub / GitLab / Bitbucket, leaving explanations and fix suggestions; reads existing comments to avoid duplicates; Default / High / Custom effort levels; usage-based billing; trigger manually with cursor review / bugbot run, or run /review locally before pushing. Per the official changelog (2026-06-10), Bugbot is now powered by Composer 2.5, averaging ~90-second reviews. As of 2026-07-10. It's now powered by Cursor's own Composer 2.5 — the official figure is about 90 seconds per review — and you can run /review locally before pushing; open the PR and Bugbot recognizes the already-reviewed diff instead of reviewing it twice.4注 4Note 4Cursor Docs · Bugbot —— 在 GitHub / GitLab / Bitbucket 的 PR 上自动评审 diff，留下解释与修复建议；会读已有评论避免重复；效果档 Default / High / Custom；usage-based 计费；可用 cursor review / bugbot run 手动触发，推送前可在本地 /review 先跑。官方 changelog（2026-06-10）称 Bugbot 现由 Composer 2.5 驱动、平均评审约 90 秒。截至 2026-07-10。Cursor Docs · Bugbot — automatically reviews PR diffs on GitHub / GitLab / Bitbucket, leaving explanations and fix suggestions; reads existing comments to avoid duplicates; Default / High / Custom effort levels; usage-based billing; trigger manually with cursor review / bugbot run, or run /review locally before pushing. Per the official changelog (2026-06-10), Bugbot is now powered by Composer 2.5, averaging ~90-second reviews. As of 2026-07-10. It's one shape of 'sent-out' work: you stop doing the mechanical scan on every PR yourself.诚实的边界：它抓的是机械问题（bug、安全、规范），不替你判断设计意图和取舍。它替你看「这改得对不对」，你还得看「该不该这么设计」—— 意图层的 review 仍然是你的。把它当成给你省下第一遍粗筛的工具，不是替你拍板的评审者。The honest boundary: it catches mechanical issues (bugs, security, conventions), not your design intent and trade-offs. It checks 'is this change correct'; you still check 'should it be designed this way' — intent-level review stays yours. Treat it as saving you the first coarse pass, not as the reviewer who decides.
— IV
SDK：从用 agent 到编排 agentThe SDK: From Using Agents to Orchestrating Them.
CLI 把 agent 放进脚本，SDK 把 agent 放进程序。The CLI puts the agent in a script; the SDK puts it in a program.@cursor/sdk（TypeScript / Python）让你在自己的代码里调用同一个 agent：local runtime 跑在你的进程里、文件走本地磁盘，cloud runtime 跑在隔离 VM 里，一把 API key 通用。6注 6Note 6Cursor Docs · SDK（TypeScript / Python）—— @cursor/sdk 让你在自己的代码里调用 Cursor 的 agent：同一个 agent，local（在你的进程里、文件走本地磁盘）或 cloud（隔离 VM）两种 runtime，一把 CURSOR_API_KEY 通用；可把自己的函数注册成工具（custom tools）、headless 跑时接 auto-review 分类器管权限、subagent 可嵌套；官方给的起点是 CI 自动修复、bug 分诊、code review、嵌入产品的 agent 与编排器。截至 2026-07-10。Cursor Docs · SDK (TypeScript / Python) — @cursor/sdk lets you call Cursor's agent from your own code: the same agent, in a local runtime (in your process, files from disk) or cloud runtime (isolated VM), one CURSOR_API_KEY for both; register your own functions as custom tools, route headless runs through the auto-review classifier for permissions, nest subagents; the official starting points are CI auto-fix bots, bug triage, code-review passes, embedded in-product agents, and orchestrators. As of 2026-07-10.比 CLI 多出来的是「编排」的那层：把你自己的函数注册成它能调的工具（custom tools）；headless 跑时不必在「全放行」和「跑不动」之间二选一 —— 接 auto-review 分类器，用自然语言写哪些调用放行、哪些拦下；subagent 还能嵌套，审稿的可以再派一个写测试的。官方给的起点很说明定位：CI 自动修复机器人、bug 分诊、code review、嵌进你自己产品里的 agent。连云端 agent 的会话本身都能挂 hook（3.11 起有 prompt 提交前、响应后、subagent 启动等节点），拿来搭自纠错的循环。@cursor/sdk (TypeScript / Python) lets you call the same agent from your own code: the local runtime runs in your process with files from disk, the cloud runtime in an isolated VM, one API key for both.6注 6Note 6Cursor Docs · SDK（TypeScript / Python）—— @cursor/sdk 让你在自己的代码里调用 Cursor 的 agent：同一个 agent，local（在你的进程里、文件走本地磁盘）或 cloud（隔离 VM）两种 runtime，一把 CURSOR_API_KEY 通用；可把自己的函数注册成工具（custom tools）、headless 跑时接 auto-review 分类器管权限、subagent 可嵌套；官方给的起点是 CI 自动修复、bug 分诊、code review、嵌入产品的 agent 与编排器。截至 2026-07-10。Cursor Docs · SDK (TypeScript / Python) — @cursor/sdk lets you call Cursor's agent from your own code: the same agent, in a local runtime (in your process, files from disk) or cloud runtime (isolated VM), one CURSOR_API_KEY for both; register your own functions as custom tools, route headless runs through the auto-review classifier for permissions, nest subagents; the official starting points are CI auto-fix bots, bug triage, code-review passes, embedded in-product agents, and orchestrators. As of 2026-07-10. What it adds over the CLI is the orchestration layer: register your own functions as tools it can call (custom tools); in headless runs, skip the choice between 'allow everything' and 'nothing runs' — wire in the auto-review classifier and describe in plain English which calls pass and which hold; subagents even nest, so a reviewer can delegate to a test-writer. The official starting points say what it's for: CI auto-fix bots, bug triage, code review, agents embedded in your own product. Even a cloud agent's conversation can carry hooks now (since 3.11: before prompt submission, after responses, on subagent start), for building self-correcting loops.判断：个人 builder 大多停在 CLI 就够 —— 一行命令、一条 automation，能覆盖绝大多数「派出去」的活。SDK 值得上的时刻只有一个：agent 不再是你手里的工具，而要成为你产品或流水线里的一个部件。到那一步，这一章讲的验收纪律不变，只是从「你审它的 PR」变成「你的程序审它的输出」。The call: most individual builders can stop at the CLI — one command or one automation covers nearly all 'send it out' work. The SDK earns its place at exactly one moment: when the agent stops being a tool in your hand and becomes a part of your product or pipeline. At that point the acceptance discipline of this chapter doesn't change — it just shifts from 'you review its PR' to 'your program reviews its output.'把一件长跑或定时的活派出去一次：Send one long-running or scheduled task out, once:
01
挑一件「做完通知我」的活Pick a 'done, tell me' task
一个跨度较长、验收标准清楚的任务，编码或非编码都行（如定时整理一批数据 / 跑一份报告）。A longer task with clear acceptance criteria, code or non-code (e.g. tidy a batch of data on a schedule / run a report).
02
交给 cloud agent，或写成一条 headless CLI 命令Hand it to a cloud agent, or write it as one headless CLI command
然后离开 —— 去网页 / Slack 看它的进度，而不是守着它。Then walk away — check its progress from web / Slack instead of watching it.
03
回收结果，决定它值不值得变成自动化Collect the result, decide if it's worth turning into an automation
如果它跑得稳、你以后还会重复，就设条 Automation 让它按时间或事件自己跑。If it ran reliably and you'll repeat it, set up an Automation to fire it on a schedule or event.
陪你做的活留在眼前，
做完通知的活派出去.
Work-with-me stays in front of you;
done-tell-me gets sent out.
Aklman Library
— 讨论Discussion
讨论Discussion.
评论区初始化中…Initializing comments…
01 / 01
没有匹配结果No matches.
换个关键词，或按 Esc 回到页面Try another keyword, or press Esc to return