边界 · 画清哪些活不该交给它The Boundary · Draw the Line on What Not to Hand It.
元工具不等于万能。命令行、文件、脚本、系统在线内;去点别的 GUI app 在线外。Warp 自己的 Computer Use 是云端沙箱里验证它写的界面,不是替你开本机应用。这章把线画清,也修正一条旧建议:远程与 SSH 场景,现在多了 Warp Agent CLI 这个选项。A meta-tool isn't omnipotent. Command line, files, scripts, system: inside the line; clicking other GUI apps: outside it. Warp's own Computer Use verifies the UI it wrote in a cloud sandbox — it doesn't drive your local apps. This chapter draws the line, and corrects one old piece of advice: for remote and SSH, there's now the Warp Agent CLI.
前面几章都在说 Warp 能干多少活。这一章反过来:说它不该干的。元工具不等于万能 —— Warp 的 agent 强在命令行、文件、脚本、系统这一侧;一旦活儿要去点别的 GUI app,那是另一类工具的事。这章也把一条旧建议改掉:以前说「远程和 SSH 场景回普通终端」,现在多了 Warp Agent CLI 这个选项。线画清楚,比再多一个功能重要。The last few chapters were about how much Warp can do. This one is the reverse: what it shouldn't. A meta-tool isn't omnipotent — Warp's agent is strong on the command-line, file, script, and system side; the moment a job means clicking some other GUI app, that's another class of tool. This chapter also corrects one old piece of advice: 'remote and SSH mean a plain terminal' — there's now the Warp Agent CLI option. Drawing the line clearly matters more than one more feature.
— I
元工具不等于万能A Meta-Tool Isn't Omnipotent.
命令、文件、代码库、交互式终端程序、MCP —— 这些都在线内,agent 在你真实的终端会话里动手。4注 4Note 4Warp Docs · Agents in Warp —— Warp agent 的能力围绕命令、文件、代码库、终端程序与 MCP;本地 agent 在你的真实终端会话里动手。Warp Docs · Agents in Warp — the Warp agent's abilities center on commands, files, the codebase, terminal programs, and MCP; local agents act inside your real terminal session.但要在你本机的设计软件里调图层、在某个网页后台里一页页点、在桌面邮件客户端里收发,那不是它的能力范围。把它当成「终端里的元工具」,而不是「替你操作电脑的万能助手」—— 这个定位差一格,期待就不会落空。判断很简单:这件事能不能用命令、文件、脚本表达?能,在线内;只能靠点一个非终端的图形界面,出界。Commands, files, the codebase, interactive terminal programs, MCP — all inside the line, with the agent acting in your real terminal session.4注 4Note 4Warp Docs · Agents in Warp —— Warp agent 的能力围绕命令、文件、代码库、终端程序与 MCP;本地 agent 在你的真实终端会话里动手。Warp Docs · Agents in Warp — the Warp agent's abilities center on commands, files, the codebase, terminal programs, and MCP; local agents act inside your real terminal session. But adjusting layers in your local design software, clicking through a web admin panel page by page, sending mail in a desktop client — that's outside its range. Treat it as 'the meta-tool inside the terminal,' not 'an all-purpose assistant that operates your computer' — that one notch of positioning keeps your expectations from falling through. The test is simple: can the task be expressed as commands, files, scripts? If yes, inside the line; if it only works by clicking a non-terminal GUI, outside it.
留在 WarpKeep in Warp
不该留在 Warp 本地会话Should not stay in local Warp session
在你本机替你点别的 appclick other apps on your local machine
— II
它的 Computer Use 是验证,不是替你点Its Computer Use Is Verification, Not Clicking for You.
Warp 确实有 Computer Use —— 但它是沙箱里给 agent 验自己的活,碰不到你本机。它只在 Warp 的沙箱云端环境跑,和你的本机、凭据隔离,用途是让 agent 在反馈闭环里验证自己改出来的界面 —— 它改了前端,自己开个浏览器看渲染对不对、表单校验过不过,完事留一段剪掉空闲时间的标注录像,可以直接附到 PR 上当验收证据。1注 1Note 1Warp Docs · Computer Use —— 截至 2026-08-06:仍只在 Warp 沙箱云端环境,与本机和凭据隔离;但默认值变了:Warp 内置 harness 的云端 run 默认开启(服务端默认),从 Warp app 发起的 run 跟随 app 设置(Settings > Agents > Warp Agent > Experimental > Computer use in Cloud Agents,app 里默认关),第三方 harness 默认关;oz 有 --computer-use / --no-computer-use,API 有 config.computer_use_enabled;带 Chromium + 脚本化 Playwright;产物是标注过的录像 artifact(空闲时间剪掉,可附到 PR);目前只支持 Anthropic Claude 模型(自动选)。Warp Docs · Computer Use — as of 2026-08-06: still only in Warp's sandboxed cloud environment, isolated from your machine and credentials; but the defaults changed — cloud runs on Warp's built-in harness have it on by default (server default), runs started from the Warp app follow the app setting (Settings > Agents > Warp Agent > Experimental > Computer use in Cloud Agents, off by default in the app), third-party harnesses default off; oz has --computer-use / --no-computer-use, the API has config.computer_use_enabled; bundled Chromium + scripted Playwright; the artifact is an annotated video recording (idle time trimmed, attachable to PRs); currently Anthropic Claude models only (auto-selected).Warp does have Computer Use — but it's a sandbox for the agent to check its own work, and it can't touch your machine. It runs only in Warp's sandboxed cloud environment, isolated from your machine and credentials, and its use is letting the agent verify its own UI changes in a feedback loop — it changed the frontend, so it opens a browser to check rendering and form validation, leaving an annotated recording with idle time trimmed that can attach to the PR as review evidence.1注 1Note 1Warp Docs · Computer Use —— 截至 2026-08-06:仍只在 Warp 沙箱云端环境,与本机和凭据隔离;但默认值变了:Warp 内置 harness 的云端 run 默认开启(服务端默认),从 Warp app 发起的 run 跟随 app 设置(Settings > Agents > Warp Agent > Experimental > Computer use in Cloud Agents,app 里默认关),第三方 harness 默认关;oz 有 --computer-use / --no-computer-use,API 有 config.computer_use_enabled;带 Chromium + 脚本化 Playwright;产物是标注过的录像 artifact(空闲时间剪掉,可附到 PR);目前只支持 Anthropic Claude 模型(自动选)。Warp Docs · Computer Use — as of 2026-08-06: still only in Warp's sandboxed cloud environment, isolated from your machine and credentials; but the defaults changed — cloud runs on Warp's built-in harness have it on by default (server default), runs started from the Warp app follow the app setting (Settings > Agents > Warp Agent > Experimental > Computer use in Cloud Agents, off by default in the app), third-party harnesses default off; oz has --computer-use / --no-computer-use, the API has config.computer_use_enabled; bundled Chromium + scripted Playwright; the artifact is an annotated video recording (idle time trimmed, attachable to PRs); currently Anthropic Claude models only (auto-selected).两处默认值 2026 年夏变了,别把老印象当事实:Warp 内置 harness 的云端 run,Computer Use 默认开(服务端默认);从 app 发起的 run 跟随 app 设置,默认关;第三方 harness 默认关。目前只支持 Anthropic Claude 模型,自动选,不用你挑。1注 1Note 1Warp Docs · Computer Use —— 截至 2026-08-06:仍只在 Warp 沙箱云端环境,与本机和凭据隔离;但默认值变了:Warp 内置 harness 的云端 run 默认开启(服务端默认),从 Warp app 发起的 run 跟随 app 设置(Settings > Agents > Warp Agent > Experimental > Computer use in Cloud Agents,app 里默认关),第三方 harness 默认关;oz 有 --computer-use / --no-computer-use,API 有 config.computer_use_enabled;带 Chromium + 脚本化 Playwright;产物是标注过的录像 artifact(空闲时间剪掉,可附到 PR);目前只支持 Anthropic Claude 模型(自动选)。Warp Docs · Computer Use — as of 2026-08-06: still only in Warp's sandboxed cloud environment, isolated from your machine and credentials; but the defaults changed — cloud runs on Warp's built-in harness have it on by default (server default), runs started from the Warp app follow the app setting (Settings > Agents > Warp Agent > Experimental > Computer use in Cloud Agents, off by default in the app), third-party harnesses default off; oz has --computer-use / --no-computer-use, the API has config.computer_use_enabled; bundled Chromium + scripted Playwright; the artifact is an annotated video recording (idle time trimmed, attachable to PRs); currently Anthropic Claude models only (auto-selected).线没变:沙箱里验证自己写的(Warp) vs 在你机器上替你动手(Claude in Chrome、Operator 那类)。要后者,就该开后者,别指望 Warp 这个沙箱伸到你的桌面上来。Two defaults changed in summer 2026 — don't treat old impressions as fact: on cloud runs with Warp's built-in harness, Computer Use is on by default (server default); runs started from the app follow the app setting, which is off; third-party harnesses default off. It currently supports Anthropic Claude models only, auto-selected — no picking for you.1注 1Note 1Warp Docs · Computer Use —— 截至 2026-08-06:仍只在 Warp 沙箱云端环境,与本机和凭据隔离;但默认值变了:Warp 内置 harness 的云端 run 默认开启(服务端默认),从 Warp app 发起的 run 跟随 app 设置(Settings > Agents > Warp Agent > Experimental > Computer use in Cloud Agents,app 里默认关),第三方 harness 默认关;oz 有 --computer-use / --no-computer-use,API 有 config.computer_use_enabled;带 Chromium + 脚本化 Playwright;产物是标注过的录像 artifact(空闲时间剪掉,可附到 PR);目前只支持 Anthropic Claude 模型(自动选)。Warp Docs · Computer Use — as of 2026-08-06: still only in Warp's sandboxed cloud environment, isolated from your machine and credentials; but the defaults changed — cloud runs on Warp's built-in harness have it on by default (server default), runs started from the Warp app follow the app setting (Settings > Agents > Warp Agent > Experimental > Computer use in Cloud Agents, off by default in the app), third-party harnesses default off; oz has --computer-use / --no-computer-use, the API has config.computer_use_enabled; bundled Chromium + scripted Playwright; the artifact is an annotated video recording (idle time trimmed, attachable to PRs); currently Anthropic Claude models only (auto-selected). The line hasn't moved: verifying what it wrote, in a sandbox (Warp) vs acting for you, on your machine (Claude in Chrome, Operator, that class). For the latter, open the latter — don't expect Warp's sandbox to reach out onto your desktop.
何时普通终端或专用工具更好When a Plain Terminal or Dedicated Tool Is Better.
有些活留在 Warp 里反而更费 —— 该出手时就切出去。Some work costs more kept inside Warp — when it's time, step out.出界规则更新到四条半:离线、或锁死到装不了任何东西的环境,普通终端仍最稳;远程与 SSH 的旧建议作废 —— 以前说「回普通终端」,现在有 Warp Agent CLI:warp 二进制把同一个 Warp Agent 带进任何终端,包括 SSH 会话,同账户、同 rules、同 skills,会话还能同步回账户、顺手 handoff 上云;服务器上跑一次 /tui-migrate-setup 还能把 app 里兼容的设置带过去(凭据不拷)。3注 3Note 3Warp Docs · Warp Agent CLI —— 截至 2026-08-06:warp 二进制(2026-07-23 由 warp-tui 改名)把 Warp Agent 带进任何终端,包括 SSH 会话;与 app 同一个 harness、账户、计划、rules、skills;自带 PTY(长跑与全屏程序可用);流式 transcript 带 diff / 工具调用 / 权限卡;首次跑 device-authorization 登录,CI 用 WARP_API_KEY / --api-key;会话持久并同步到账户;内置云端 handoff;认共享路径的 AGENTS.md / skills / MCP;Drive 里存的 Prompts 可用,其他 Drive 对象不行;/tui-migrate-setup 能拷兼容的 app 设置(凭据不拷);macOS(Apple Silicon + Intel)、Linux / Windows(x64 + Arm64)。Warp Docs · Warp Agent CLI — as of 2026-08-06: the warp binary (renamed from warp-tui on 2026-07-23) brings the Warp Agent to any terminal, including over SSH; same harness, account, plan, rules, and skills as the app; its own PTY (long-running and full-screen apps work); streaming transcript with diffs / tool calls / permission cards; device-authorization login on first run, WARP_API_KEY / --api-key for CI; conversations persist and sync to the account; cloud handoff built in; picks up AGENTS.md / skills / MCP from shared paths; saved Warp Drive Prompts are available, other Drive object types are not; /tui-migrate-setup copies compatible app settings (credentials never copied); macOS (Apple Silicon + Intel), Linux / Windows (x64 + Arm64).稳定重复的流水线,写成脚本挂 cron 比每次叫 agent 强;要驱动 GUI,用 Computer Use 那类工具;大段、跨多文件、靠编辑器来回改的活,切到 Cursor —— 终端侧和编辑器侧本就是一对元工具,下一章讲分工。2注 2Note 2Warp Docs · Full Terminal Use —— agent 在 PTY 里与交互式终端程序协作(psql、vim、gdb、REPL、dev server);这是 Warp agent 的主场,与「驱动桌面 GUI」是两回事。Warp Docs · Full Terminal Use — the agent works with interactive terminal apps inside the PTY (psql, vim, gdb, REPLs, dev servers); this is the Warp agent's home turf, distinct from 'driving a desktop GUI.'知道什么时候不用它,和知道怎么用它一样,是判断力。The out-of-bounds rules are now four and a half: offline, or an environment so locked down nothing can be installed — a plain terminal is still steadiest; the old remote-and-SSH advice is retired — it used to be 'go back to a plain terminal,' but now there's the Warp Agent CLI: the warp binary brings the same Warp Agent to any terminal, SSH sessions included, with the same account, rules, and skills, conversations syncing back to your account with cloud handoff built in; one /tui-migrate-setup on the server carries over compatible app settings (never credentials).3注 3Note 3Warp Docs · Warp Agent CLI —— 截至 2026-08-06:warp 二进制(2026-07-23 由 warp-tui 改名)把 Warp Agent 带进任何终端,包括 SSH 会话;与 app 同一个 harness、账户、计划、rules、skills;自带 PTY(长跑与全屏程序可用);流式 transcript 带 diff / 工具调用 / 权限卡;首次跑 device-authorization 登录,CI 用 WARP_API_KEY / --api-key;会话持久并同步到账户;内置云端 handoff;认共享路径的 AGENTS.md / skills / MCP;Drive 里存的 Prompts 可用,其他 Drive 对象不行;/tui-migrate-setup 能拷兼容的 app 设置(凭据不拷);macOS(Apple Silicon + Intel)、Linux / Windows(x64 + Arm64)。Warp Docs · Warp Agent CLI — as of 2026-08-06: the warp binary (renamed from warp-tui on 2026-07-23) brings the Warp Agent to any terminal, including over SSH; same harness, account, plan, rules, and skills as the app; its own PTY (long-running and full-screen apps work); streaming transcript with diffs / tool calls / permission cards; device-authorization login on first run, WARP_API_KEY / --api-key for CI; conversations persist and sync to the account; cloud handoff built in; picks up AGENTS.md / skills / MCP from shared paths; saved Warp Drive Prompts are available, other Drive object types are not; /tui-migrate-setup copies compatible app settings (credentials never copied); macOS (Apple Silicon + Intel), Linux / Windows (x64 + Arm64). A stable repeated pipeline — a script on cron beats calling the agent each time; driving a GUI — use a Computer Use class tool; large, multi-file, editor-driven rewriting — switch to Cursor, since the terminal side and the editor side are a pair of meta-tools, and the next chapter covers the split.2注 2Note 2Warp Docs · Full Terminal Use —— agent 在 PTY 里与交互式终端程序协作(psql、vim、gdb、REPL、dev server);这是 Warp agent 的主场,与「驱动桌面 GUI」是两回事。Warp Docs · Full Terminal Use — the agent works with interactive terminal apps inside the PTY (psql, vim, gdb, REPLs, dev servers); this is the Warp agent's home turf, distinct from 'driving a desktop GUI.' Knowing when not to use it is as much judgment as knowing how to use it.
提示词Prompt留在 Warp 还是切走Stay in Warp or switch away
判断这个任务该留在 Warp 还是切走:
任务:<描述>
它能否用命令/文件/脚本/日志表达:<是/否>
是否需要 GUI 点击、真实账号、付款、发送、权限变更:<是/否>
是否可 dry-run、小样本、回滚:<是/否>
涉及数据:<公开/内部/客户/生产/密钥>
请输出:留在 Warp / Warp Agent CLI / 普通终端 / Cursor / 浏览器 agent / 手动,并说明边界。Decide whether this task should stay in Warp or switch away:
Task: <description>
Can it be expressed as commands/files/scripts/logs: <yes/no>
Does it require GUI clicks, real account, payment, sending, permission changes: <yes/no>
Can it dry-run, sample, roll back: <yes/no>
Data involved: <public/internal/customer/production/secrets>
Output: stay in Warp / Warp Agent CLI / plain terminal / Cursor / browser agent / manual, and explain the boundary.
判断这个任务该留在 Warp 还是切走:
任务:SSH 到生产只读备库,查慢查询 top 10 并给出索引建议
它能否用命令/文件/脚本/日志表达:是
是否需要 GUI 点击、真实账号、付款、发送、权限变更:否
是否可 dry-run、小样本、回滚:是(全程只读 SELECT)
涉及数据:生产(只读)
请输出:留在 Warp / Warp Agent CLI / 普通终端 / Cursor / 浏览器 agent / 手动,并说明边界。Decide whether this task should stay in Warp or switch away:
Task: SSH to the production read replica, find the top 10 slow queries, and suggest indexes
Can it be expressed as commands/files/scripts/logs: yes
Does it require GUI clicks, real account, payment, sending, permission changes: no
Can it dry-run, sample, roll back: yes (read-only SELECTs throughout)
Data involved: production (read-only)
Output: stay in Warp / Warp Agent CLI / plain terminal / Cursor / browser agent / manual, and explain the boundary.
验收这一章 —— 把边界落到你自己的清单上:Accept the chapter — put the boundary onto your own list:
需要切走的信号:任务核心是看屏幕、点按钮、跨账号状态、不可逆提交。Switch-away signals: core task is seeing the screen, clicking buttons, cross-account state, irreversible submit.
如果必须做 GUI 自动化:只用测试账号、测试数据、测试环境;最后一步提交、付款、发送必须人工确认。If GUI automation is required: test account, data, environment only; final submit / payment / send is human-confirmed.
写下 3 件你曾想(或差点)让 Warp agent 做、但它不在主场的活,并归类:普通终端 / 写脚本 / 切 Cursor / Computer Use 那类工具。Write down 3 tasks you wanted (or nearly had) the Warp agent do that weren't its home turf, classified: plain terminal / script / Cursor / Computer Use class tool.
知道远程/SSH 的新答案:装一个 warp CLI,同一个 agent 跟着你进服务器;装不了的地方才回普通终端。Know the new remote/SSH answer: install the warp CLI and the same agent follows you into the server; only where nothing can be installed do you fall back to a plain terminal.
知道 Computer Use 的默认值:云端内置 harness 默认开,app 发起的默认关;它在沙箱里验证,不替你点本机。Know Computer Use's defaults: on by default for cloud runs on the built-in harness, off for app-started runs; it verifies in a sandbox, it doesn't click your local machine.
它能验证它写的界面,
碰不到你开的应用.
It can verify the UI it wrote,
it can't touch the apps you opened.
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
