放出去 · 云端 agent 与 Oz 编排Send It Out · Cloud Agents and Oz Orchestration.
有些活不用你守着：夜里跑的维护、GitHub 上一开 PR 就触发的检查、Slack 里一句话起一个任务。Warp 的云端 agent 自己在后台跑，本地起的对话还能交接上云（反过来也能用 /continue-locally 拉回本地）；Oz 是把多个 agent 编排成一群的那层。这章讲哪些值得放上云、调度与事件触发怎么用，以及对一个人用户，多 agent 编排什么时候是过度。Some work does not need you watching: maintenance that runs overnight, a check that fires the moment a PR opens on GitHub, a task kicked off by one line in Slack. Warp's cloud agents run in the background on their own, and a conversation you started locally can hand off to the cloud (and back, via /continue-locally); Oz is the layer that orchestrates many agents into a swarm. What is worth sending up, how scheduling and event triggers work, and when multi-agent orchestration is overkill for a solo user.
到这里 agent 都在你眼前跑 —— 你看着它一步步动。但有些活不值得你守着：夜里的维护、一开 PR 就该触发的检查、Slack 里一句话起的任务。Warp 的云端 agent 自己在后台跑，本地起的对话还能交接上云；Oz 是把多个 agent 编排成一群的那层。这章讲哪些值得放上云、对一个人何时是过度，以及上云那一刻什么离开了你的机器。So far the agent has run in front of you — you watch it move step by step. But some work isn't worth watching: overnight maintenance, a check that should fire the moment a PR opens, a task kicked off by one line in Slack. Warp's cloud agents run in the background on their own, and a conversation you started locally can hand off to the cloud; Oz is the layer that orchestrates many agents into a swarm. This chapter covers what's worth sending up, when it's overkill for one person, and what leaves your machine the moment you go cloud.
— I
云端 agent：不用你守着Cloud Agents: No Need to Watch.
本地 agent 要你在场；云端 agent 在你离场时替你跑。A local agent needs you present; a cloud agent runs for you while you're away.云端 agent 自主运行，能按计划定时跑、响应 Slack / Linear / GitHub 的事件、并行执行 —— PR 一开就评审、夜里做维护、定时出报告，都不用你盯着。1注 1Note 1Warp Docs · Oz Platform overview —— 截至 2026-07-10，Cloud Agents 运行在 Oz Platform 上：trigger（schedule、Slack / GitHub / Linear 等集成、API、手动）创建任务，编排层跟踪生命周期，任务在 host（可选 environment）上执行，留下 status、metadata、transcript、outputs 的持久记录，团队可检视与分享；self-hosted execution 需要 Enterprise 计划。原 agent-platform/cloud-agents 一节已迁往 /platform/。Warp Docs · Oz Platform overview — as of 2026-07-10, Cloud Agents run on the Oz Platform: triggers (schedules, Slack / GitHub / Linear integrations, API calls, manual runs) create tasks, an orchestration layer tracks their lifecycle, tasks execute on a host (optionally in an environment), and each leaves a persistent record — status, metadata, transcript, outputs — the team can inspect and share; self-hosted execution requires an Enterprise plan. The former agent-platform/cloud-agents section has moved to /platform/.命令行起一个也简单：Cloud agents run autonomously — on a schedule, reacting to Slack / Linear / GitHub events, in parallel — so a PR gets reviewed the moment it opens, maintenance happens overnight, reports go out on time, none of it watched.1注 1Note 1Warp Docs · Oz Platform overview —— 截至 2026-07-10，Cloud Agents 运行在 Oz Platform 上：trigger（schedule、Slack / GitHub / Linear 等集成、API、手动）创建任务，编排层跟踪生命周期，任务在 host（可选 environment）上执行，留下 status、metadata、transcript、outputs 的持久记录，团队可检视与分享；self-hosted execution 需要 Enterprise 计划。原 agent-platform/cloud-agents 一节已迁往 /platform/。Warp Docs · Oz Platform overview — as of 2026-07-10, Cloud Agents run on the Oz Platform: triggers (schedules, Slack / GitHub / Linear integrations, API calls, manual runs) create tasks, an orchestration layer tracks their lifecycle, tasks execute on a host (optionally in an environment), and each leaves a persistent record — status, metadata, transcript, outputs — the team can inspect and share; self-hosted execution requires an Enterprise plan. The former agent-platform/cloud-agents section has moved to /platform/. Kicking one off from the command line is simple too:本机只做了轻量 CLI 核验：Warp app bundle 里带 oz，版本为 Oz v0.2026.07.01.09.21.stable_01，oz --help 列出 agent、environment、mcp、schedule、secret、memory-store 等命令；「cloud agents 的 orchestration platform」是官方 docs 给 Oz 的定义。3注 3Note 3本机轻测 —— 2026-07-10，本机 /Applications/Warp.app/Contents/Resources/bin/oz --version 输出 Oz v0.2026.07.01.09.21.stable_01；oz --help 列出 agent / environment / mcp / run / model / memory-store / memory / integration / schedule / secret / api-key 等命令；「cloud agents 的 orchestration platform」是 docs 首页给 Oz 的定义。Local light check — on 2026-07-10, /Applications/Warp.app/Contents/Resources/bin/oz --version returned Oz v0.2026.07.01.09.21.stable_01; oz --help lists commands including agent, environment, mcp, run, model, memory-store, memory, integration, schedule, secret, and api-key; 'the orchestration platform for cloud agents' is how the docs home defines Oz.没有账户权限的 Cloud Agent / Oz Web App / Computer Use，本书按官方 docs 写，不冒充实测。Local validation here is intentionally light: the Warp app bundle includes oz, version Oz v0.2026.07.01.09.21.stable_01, and oz --help lists agent, environment, mcp, schedule, secret, memory-store, and more; 'the orchestration platform for cloud agents' is how the official docs define Oz.3注 3Note 3本机轻测 —— 2026-07-10，本机 /Applications/Warp.app/Contents/Resources/bin/oz --version 输出 Oz v0.2026.07.01.09.21.stable_01；oz --help 列出 agent / environment / mcp / run / model / memory-store / memory / integration / schedule / secret / api-key 等命令；「cloud agents 的 orchestration platform」是 docs 首页给 Oz 的定义。Local light check — on 2026-07-10, /Applications/Warp.app/Contents/Resources/bin/oz --version returned Oz v0.2026.07.01.09.21.stable_01; oz --help lists commands including agent, environment, mcp, run, model, memory-store, memory, integration, schedule, secret, and api-key; 'the orchestration platform for cloud agents' is how the docs home defines Oz. Cloud Agent / Oz Web App / Computer Use flows that need account access are written from official docs, not pretended as local tests.
任务：<后台要完成什么>
触发：<手动 / schedule / Slack / GitHub / Linear / webhook>
环境：<repo / image / startup commands / secrets>
边界：只允许 <读/写范围>；禁止 <生产写入/删除/deploy/secret 输出>
验收：<测试命令 / PR / 报告 / artifact / run transcript>
通知：完成后把摘要发到 <人/频道/issue>
失败：失败时停止，保留 logs，不自动重试破坏性命令。Task: <what should finish in the background>
Trigger: <manual / schedule / Slack / GitHub / Linear / webhook>
Environment: <repo / image / startup commands / secrets>
Boundary: allow only <read/write scope>; forbid <production write/delete/deploy/secret output>
Acceptance: <test command / PR / report / artifact / run transcript>
Notify: send summary to <person/channel/issue> when complete
Failure: stop on failure, preserve logs, do not auto-retry destructive commands.
任务：每周一早上跑一次依赖审计，找出 npm 里有已知漏洞或落后两个大版本的包
触发：schedule，每周一 08:00
环境：repo github.com/acme/web，node:22 image，启动跑 npm ci
边界：只允许读 package.json / package-lock.json 和跑 npm audit；禁止生产写入、删除、deploy、secret 输出
验收：生成一份按严重度排序的报告 artifact，附升级建议命令
通知：完成后把摘要发到 #dependencies 频道
失败：失败时停止，保留 logs，不自动重试破坏性命令。Task: run a dependency audit every Monday morning to find npm packages with known vulnerabilities or two major versions behind
Trigger: schedule, every Monday 08:00
Environment: repo github.com/acme/web, node:22 image, run npm ci on startup
Boundary: allow only reading package.json / package-lock.json and running npm audit; forbid production writes, deletes, deploys, secret output
Acceptance: produce a report artifact sorted by severity, with suggested upgrade commands
Notify: send summary to the #dependencies channel when complete
Failure: stop on failure, preserve logs, do not auto-retry destructive commands.
— II
本地起，云端接Start Local, Hand to Cloud.
你不必一开始就决定本地还是云端 —— 一段对话可以跑着跑着就交接上去。You don't have to decide local or cloud up front — a conversation can hand off mid-run.本地起的 agent 对话能交接上云继续，反过来用 /continue-locally 拉回本地，/conversations 访问云端那些。2注 2Note 2Warp Docs · Getting started with Warp and Oz —— Oz 是 Warp 智能功能背后的 cloud-agent orchestration platform；Cloud Agents 可跑在 Warp infrastructure 或自有 infrastructure；Warp 与 Oz 共享 Drive、Rules、MCP 等上下文。Warp Docs · Getting started with Warp and Oz — Oz is the cloud-agent orchestration platform behind Warp's intelligent features; Cloud Agents can run on Warp infrastructure or your own; Warp and Oz share context such as Drive, Rules, and MCP.Oz 是这背后的编排平台（2026-02 上线），把单个 agent 扩成可管理、可并行的一群。判断线很清楚：要你来回确认、需要你判断的活，留本地；明确、可验证、能跑很久的活，交接上云让它自己走。A locally started agent conversation can hand off to the cloud to continue, and back via /continue-locally, with /conversations reaching the cloud ones.2注 2Note 2Warp Docs · Getting started with Warp and Oz —— Oz 是 Warp 智能功能背后的 cloud-agent orchestration platform；Cloud Agents 可跑在 Warp infrastructure 或自有 infrastructure；Warp 与 Oz 共享 Drive、Rules、MCP 等上下文。Warp Docs · Getting started with Warp and Oz — Oz is the cloud-agent orchestration platform behind Warp's intelligent features; Cloud Agents can run on Warp infrastructure or your own; Warp and Oz share context such as Drive, Rules, and MCP. Oz is the orchestration platform behind this (launched 2026-02), scaling a single agent into a manageable, parallel swarm. The line is clear: work that needs your back-and-forth and judgment stays local; work that's well-defined, verifiable, and long-running hands off to the cloud to walk on its own.
选择Choice
适合Best for
不要用在Avoid for
本地 AgentLocal Agent
你要实时审命令、看输出、随时改方向You need to review commands, watch output, steer in real time
长时间等待、并行跑很多任务Long waits, many parallel tasks
Cloud AgentCloud Agent
明确、可验收、能后台跑、触发清楚Clear, reviewable, background task with a defined trigger
需要频繁人工判断、敏感本地凭据Frequent human judgment, sensitive local credentials
Oz orchestration
多 agent 并行、团队监督、run 可观察性、事件/计划自动化Multi-agent parallelism, team oversight, run observability, event/schedule automation
一个人临时跑一条命令A solo one-off command
— III
单人何时过度，数据走到哪When It's Overkill, and Where the Data Goes.
对一个人，有用的是一两个定时云端 agent；整套 swarm 编排是团队的答案。上云之前，先看清什么离开了你的机器。For one person, the useful part is one or two scheduled cloud agents; the full swarm orchestration is a team's answer. Before you go cloud, see what leaves your machine.多 agent 编排是给团队规模设计的 —— 单人用一两个定时或事件触发的云端 agent 就够，再往上的 swarm 大多是过度，还多烧 credits 和并发。更要算的是数据和账单：Cloud Agents 与 integrations 运行在 Oz Platform 的云端控制面上，按 credits 计费 —— 截至 2026-07-10，credits 分三桶（AI 推理、compute 沙箱、platform 编排层），同一池扣；每个 cloud run 至少消耗 compute 和 platform 两桶，用 Warp 托管的模型再加 AI 那桶，而你本机配的 BYOK 云端拿不到。4注 4Note 4Warp Docs · Platform credits —— 截至 2026-07-10，Warp 把 credits 分三桶：AI（推理）、compute（云端沙箱）、platform（编排、集成、观测），同一额度池扣减；每个 cloud agent run 都消耗 platform credits（不论 harness 是 Warp Agent、Claude Code 还是 Codex、推理来自谁）；cloud run 用 Warp 托管算力时消耗 compute credits，本地 run 用你自己的机器、不耗 compute。Warp Docs · Platform credits — as of 2026-07-10, Warp meters credits in three buckets — AI (inference), compute (the cloud sandbox), platform (orchestration, integrations, observability) — all drawing from one pool; every cloud agent run consumes platform credits regardless of whether the harness is Warp Agent, Claude Code, or Codex, or whose inference it uses; cloud runs on Warp-hosted compute consume compute credits, while local runs use your own machine and don't.6注 6Note 6Warp Docs · Bring Your Own API Key —— BYOK 不适用于 Cloud Agents：key 只存在你设备上，云端 run 拿不到，cloud runs 照常消耗 Warp credits；Enterprise 需要在云端用自己 key 的团队可由管理员配 team-managed keys（Warp 服务端保存）。Warp Docs · Bring Your Own API Key — BYOK does not apply to Cloud Agents: keys live only on your device, cloud-hosted runs can't reach them and consume Warp credits as usual; Enterprise teams that need their own keys in the cloud can have admins configure team-managed keys, stored server-side by Warp.Oz 还能把 Claude Code、Codex 当云端 harness 跑（pricing 页标 beta）：推理由你提供的凭据记 provider 的账，Warp 仍计 compute 与 platform credits —— 你已付费的 CLI agent 订阅上云能接着用，但不是免费。5注 5Note 5Warp Docs · Harnesses in Oz —— Oz 可把第三方 harness（Claude Code、Codex）当云端 agent 跑，与 Warp Agent 共享 triggers、environments、secrets 与观测；计费上，两者用你提供的凭据直接调用 provider、推理记你的 provider 账，Warp 按 compute credits（沙箱）与 platform credits（编排层）计量。Warp pricing 页把「Use any harness in the cloud」标为 beta。Warp Docs · Harnesses in Oz — Oz can run third-party harnesses (Claude Code, Codex) as cloud agents sharing Warp Agent's triggers, environments, secrets, and observability; for billing, both call the provider directly with credentials you supply so inference lands on your provider account, while Warp meters compute credits (the sandbox) and platform credits (the orchestration layer). Warp's pricing page labels 'Use any harness in the cloud' as beta.官方写到 cloud agents 可跑在 Warp infrastructure 或自有 infrastructure；但 self-hosted execution 需要 Enterprise 计划，Enterprise / team 维度的治理不要写成个人功能。2注 2Note 2Warp Docs · Getting started with Warp and Oz —— Oz 是 Warp 智能功能背后的 cloud-agent orchestration platform；Cloud Agents 可跑在 Warp infrastructure 或自有 infrastructure；Warp 与 Oz 共享 Drive、Rules、MCP 等上下文。Warp Docs · Getting started with Warp and Oz — Oz is the cloud-agent orchestration platform behind Warp's intelligent features; Cloud Agents can run on Warp infrastructure or your own; Warp and Oz share context such as Drive, Rules, and MCP.1注 1Note 1Warp Docs · Oz Platform overview —— 截至 2026-07-10，Cloud Agents 运行在 Oz Platform 上：trigger（schedule、Slack / GitHub / Linear 等集成、API、手动）创建任务，编排层跟踪生命周期，任务在 host（可选 environment）上执行，留下 status、metadata、transcript、outputs 的持久记录，团队可检视与分享；self-hosted execution 需要 Enterprise 计划。原 agent-platform/cloud-agents 一节已迁往 /platform/。Warp Docs · Oz Platform overview — as of 2026-07-10, Cloud Agents run on the Oz Platform: triggers (schedules, Slack / GitHub / Linear integrations, API calls, manual runs) create tasks, an orchestration layer tracks their lifecycle, tasks execute on a host (optionally in an environment), and each leaves a persistent record — status, metadata, transcript, outputs — the team can inspect and share; self-hosted execution requires an Enterprise plan. The former agent-platform/cloud-agents section has moved to /platform/.普通项目可以上云；碰合规、机密、客户数据、生产凭据的，先留本地或走团队治理，再谈上云。Multi-agent orchestration is built for team scale — one person is fine with one or two scheduled or event-triggered cloud agents, and the swarm beyond that is mostly overkill, burning extra credits and concurrency. The bigger ledger is data and billing: Cloud Agents and integrations run on the Oz Platform's cloud control plane and are billed in credits — as of 2026-07-10, credits come in three buckets (AI for inference, compute for the sandbox, platform for the orchestration layer), all drawn from one pool; every cloud run consumes at least the compute and platform buckets, plus AI when a Warp-managed model runs the inference, and the BYOK you configured locally is out of the cloud's reach.4注 4Note 4Warp Docs · Platform credits —— 截至 2026-07-10，Warp 把 credits 分三桶：AI（推理）、compute（云端沙箱）、platform（编排、集成、观测），同一额度池扣减；每个 cloud agent run 都消耗 platform credits（不论 harness 是 Warp Agent、Claude Code 还是 Codex、推理来自谁）；cloud run 用 Warp 托管算力时消耗 compute credits，本地 run 用你自己的机器、不耗 compute。Warp Docs · Platform credits — as of 2026-07-10, Warp meters credits in three buckets — AI (inference), compute (the cloud sandbox), platform (orchestration, integrations, observability) — all drawing from one pool; every cloud agent run consumes platform credits regardless of whether the harness is Warp Agent, Claude Code, or Codex, or whose inference it uses; cloud runs on Warp-hosted compute consume compute credits, while local runs use your own machine and don't.6注 6Note 6Warp Docs · Bring Your Own API Key —— BYOK 不适用于 Cloud Agents：key 只存在你设备上，云端 run 拿不到，cloud runs 照常消耗 Warp credits；Enterprise 需要在云端用自己 key 的团队可由管理员配 team-managed keys（Warp 服务端保存）。Warp Docs · Bring Your Own API Key — BYOK does not apply to Cloud Agents: keys live only on your device, cloud-hosted runs can't reach them and consume Warp credits as usual; Enterprise teams that need their own keys in the cloud can have admins configure team-managed keys, stored server-side by Warp. Oz can also run Claude Code or Codex as a cloud harness (the pricing page labels this beta): inference lands on your provider account via credentials you supply, while Warp still meters compute and platform credits — the CLI agent subscription you already pay for keeps working in the cloud, but not for free.5注 5Note 5Warp Docs · Harnesses in Oz —— Oz 可把第三方 harness（Claude Code、Codex）当云端 agent 跑，与 Warp Agent 共享 triggers、environments、secrets 与观测；计费上，两者用你提供的凭据直接调用 provider、推理记你的 provider 账，Warp 按 compute credits（沙箱）与 platform credits（编排层）计量。Warp pricing 页把「Use any harness in the cloud」标为 beta。Warp Docs · Harnesses in Oz — Oz can run third-party harnesses (Claude Code, Codex) as cloud agents sharing Warp Agent's triggers, environments, secrets, and observability; for billing, both call the provider directly with credentials you supply so inference lands on your provider account, while Warp meters compute credits (the sandbox) and platform credits (the orchestration layer). Warp's pricing page labels 'Use any harness in the cloud' as beta. Official docs say cloud agents can run on Warp infrastructure or your own, but self-hosted execution requires an Enterprise plan; Enterprise/team governance should not be written as a solo-user feature.2注 2Note 2Warp Docs · Getting started with Warp and Oz —— Oz 是 Warp 智能功能背后的 cloud-agent orchestration platform；Cloud Agents 可跑在 Warp infrastructure 或自有 infrastructure；Warp 与 Oz 共享 Drive、Rules、MCP 等上下文。Warp Docs · Getting started with Warp and Oz — Oz is the cloud-agent orchestration platform behind Warp's intelligent features; Cloud Agents can run on Warp infrastructure or your own; Warp and Oz share context such as Drive, Rules, and MCP.1注 1Note 1Warp Docs · Oz Platform overview —— 截至 2026-07-10，Cloud Agents 运行在 Oz Platform 上：trigger（schedule、Slack / GitHub / Linear 等集成、API、手动）创建任务，编排层跟踪生命周期，任务在 host（可选 environment）上执行，留下 status、metadata、transcript、outputs 的持久记录，团队可检视与分享；self-hosted execution 需要 Enterprise 计划。原 agent-platform/cloud-agents 一节已迁往 /platform/。Warp Docs · Oz Platform overview — as of 2026-07-10, Cloud Agents run on the Oz Platform: triggers (schedules, Slack / GitHub / Linear integrations, API calls, manual runs) create tasks, an orchestration layer tracks their lifecycle, tasks execute on a host (optionally in an environment), and each leaves a persistent record — status, metadata, transcript, outputs — the team can inspect and share; self-hosted execution requires an Enterprise plan. The former agent-platform/cloud-agents section has moved to /platform/. Ordinary projects can go cloud; compliance, secrets, customer data, and production credentials should stay local or go through team governance before cloud.
云端 agent 验收：每个 run 必须留下可审的持久记录 —— 状态、元数据、transcript、outputs。Cloud Agent review: each run must leave a reviewable persistent record — status, metadata, transcript, outputs.
团队监督规则：谁能触发、谁能看、谁能 join session、谁负责验收，要写清。Team oversight rule: who can trigger, view, join, and approve must be explicit.
无人值守禁止：生产数据库写入、删除数据、付款/发信、权限变更、deploy、secret 输出。Unattended forbidden list: production DB writes, data deletion, payments/email, permission changes, deploys, secret output.
成本规则：上云前确认 credit 余额、预算、失败重试策略。Cost rule: before cloud, confirm credit balance, budget, and failure retry policy.
把一件你定期手动做的活放上云一次：Send one chore you do manually on a schedule up to the cloud, once:
01
挑一件定期重复的活Pick one recurring chore
夜里清理、每周报表、PR 一开就跑的检查 —— 选一件你现在手动触发、且不碰机密数据的。Nightly cleanup, a weekly report, checks that should run when a PR opens — pick one you now trigger by hand and that touches no sensitive data.
02
用 run-cloud 设成定时或事件触发Set it on a schedule or event trigger with run-cloud
用 oz agent run-cloud 起，按计划或按事件自己跑，先观察一周。Launch it with oz agent run-cloud, run on a schedule or event on its own, and watch for a week.
03
写下你绝不会上云的那一件，和原因Write down the one thing you'd never send up, and why
那条「不上云」的理由 —— 多半是数据 —— 就是你云端与本地之间的边界。That 'not to the cloud' reason — usually data — is your boundary between cloud and local.
在场的活留本地，
离场的活放云端.
Work you attend stays local,
work you leave goes to the cloud.
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