— 11 / 第十一章

收束 · 一个人，一支 agent 队.

把整本书压成一句：Claude Code 的问题从来不是「它能不能写这段代码」，是「你能把多少活派出去、又能不能确认它做对了」。编辑器是它最小的样子；计划循环、记忆、Skills、subagents、MCP、hooks、headless 把它撑成一个人指挥一支 agent 队的编排台。剩下的，是知道那条边界画在哪 —— 哪些活自己盯、哪些活派出去、哪一步该切走 —— 然后让这支队伍替你把活干完，而你只做那件机器替不了的事：决定做什么，以及确认它真做成了。The whole book in one line: the question was never 'can Claude Code write this code,' but 'how much can you delegate — and can you confirm it got it right.' The editor is its smallest form; the plan loop, memory, Skills, subagents, MCP, hooks, and headless stretch it into a console where one person directs a fleet of agents. The rest is knowing where the boundary sits — what to watch, what to send out, where to switch away — and then letting the fleet finish the work, while you do the one thing a machine can't: decide what to do, and confirm it was truly done.

7 分钟 · 初稿 2026.07

MD 海报

把十章压成一句:Claude Code 的问题从来不是「它能不能写这段代码」,是「你能把多少活派出去 —— 又能不能确认它真做对了」。写代码只是它最小的样子;你付费买到的,是一个人指挥一支 agent 队的编排台。

— I

回扣:六件与两翼.

回扣一遍这台机器怎么拼起来的。循环给了节奏 —— 先批方案,再收结果(第 2 章);记忆免了重复交代(第 3 章);Skills 和输出风格把你的做法与口吻收编成资产(第 4 章);subagents 给你人手,一句话铺开一支并行的队(第 5 章);MCP 给这支队素材,工单、报错、数据自己长进上下文(第 6 章);hooks 立规矩,到点必然执行、不看模型脸色(第 7 章)。再加两翼:headless 让活离开你的终端,在 CI 里、在夜里、在你不在的地方跑完(第 8 章);验收让你敢收 —— 检查、复审、回归,凭据在手才算完(第 9 章)。四个入口,一台引擎,你的配置走到哪跟到哪。1注 1Claude Code Docs · Overview —— 终端、IDE、桌面应用与浏览器四个入口连着同一台引擎,CLAUDE.md、设置与 MCP 跨入口通用。截至 2026-07-17。Recap how the machine assembles. The loop set the cadence — approve the plan, collect the result (chapter 2); memory ended the re-briefing (chapter 3); Skills and output styles codified your procedures and your voice into assets (chapter 4); subagents gave you hands, a parallel crew fanned out in one sentence (chapter 5); MCP fed that crew its material — tickets, errors, data growing into context on their own (chapter 6); hooks laid down rules that fire on the moment, no model goodwill consulted (chapter 7). Then the two wings: headless sent the work off your terminal — into CI, into the night, into wherever you aren't (chapter 8); and review let you dare collect it — checks, re-review, regression, done only with evidence in hand (chapter 9). Four entrances, one engine, your configuration following you everywhere.1注 1Claude Code Docs · Overview —— 终端、IDE、桌面应用与浏览器四个入口连着同一台引擎,CLAUDE.md、设置与 MCP 跨入口通用。截至 2026-07-17。

— II

你只做机器替不了的两件事.

队伍成形之后,你的工作只剩两件,恰好是机器替不了的两件:决定做什么,和确认它真做成了。前者是品味和取舍,后者是这本书反复回来的那句 —— 给它一个能自己跑的检查,要证据,不要断言。2注 2Claude Code Docs · Best practices —— 验证一节的第一条:给它一个它能自己跑的检查;要证据,不要断言。截至 2026-07-17。派活的人已经太多,这两件事上的功夫,才是编排时代里你真正的手艺。Once the fleet takes shape, your work reduces to two things — precisely the two a machine can't do: deciding what to build, and confirming it was truly built. The first is taste and trade-off; the second is the line this book keeps returning to — give it a check it can run, evidence over assertions.2注 2Claude Code Docs · Best practices —— 验证一节的第一条:给它一个它能自己跑的检查;要证据,不要断言。截至 2026-07-17。 Delegators are already plentiful; craft in these two is what's actually yours in the orchestration era.

— III

边界,和一把不过期的尺.

剩下的是画边界。哪些活自己盯:不可逆的、高风险的、要品味的。哪些活派出去:可验证的、可并行的、重复的。哪一步切走:该在编辑器里陪着改的、长在 GitHub 流程里的、以终端本身为中心的 —— 入口跟着活走(第 10 章)。最后说句实话:这本书会过期,而且很快 —— Claude Code 以周为单位在变,书里每个数字都标了核对日期,过期请以官方文档为准。但这把尺子不过期:拿到任何一个新特性,还是那三问 —— 它省什么,什么活用它,怎么验收。问完这三问,你就不需要下一本教程了。One honest closing note: this book will expire, and quickly — Claude Code changes week by week, every number here carries its verification date, and the official docs outrank an expired page. But the ruler doesn't expire: hand it any new feature and ask the same three questions — what does it save, which work is it for, how do I verify it. Once you ask those three, you won't need the next tutorial.

一个人,一支 agent 队: 你派活,你验收,其余交给它们.

One person, a fleet of agents: you delegate, you verify — the rest is theirs.

Aklman Library

— 读完了 · 试试这本

技术 · 实践

Codex

多 Agent 异步编码

Chat、Work、Codex 已进同一个应用;真正的跃迁,是让一队 agent 在本地分工、送上云端,再从 diff 与 PR 里收回证据。

Aklman · 文库

Codex · 多 Agent 的异步编码工作流.

Codex 是 ChatGPT 桌面应用里独立的软件开发体验,也是 CLI、IDE 扩展、云端执行与 GitHub review 组成的一套工具链。你定义目标和验收标准,主 agent 决定怎么拆活,explorer、worker 与自定义 agent 在独立线程里调查或执行,worktree 和 cloud task 再隔离会产生改动的并行工作,最后交回可审的 diff 或 PR。这本书用十二章讲清入口、权限、AGENTS.md、config、Skills、MCP、多 Agent 配置、云端、无人值守、验收与 credits;重点不是功能数量,而是怎样让任务在你不盯着时完成,同时保留证据、边界与最终判断。Codex is the dedicated software-development experience in the ChatGPT desktop app and a toolchain spanning the CLI, IDE extension, cloud execution, and GitHub review. You define the goal and acceptance criteria; the main agent decides how to divide the work; explorer, worker, and custom agents investigate or execute in separate threads; worktrees and cloud tasks isolate parallel work that produces changes; and the result returns as a reviewable diff or PR. Across twelve chapters, this book covers surfaces, permissions, AGENTS.md, config, Skills, MCP, multi-agent configuration, cloud, unattended runs, review, and credits. The point is not feature count, but finishing work without watching while keeping evidence, boundaries, and final judgment intact.

12 章 · 118 分钟 · 2026.08

← 上一本：Warp · 一台终端形状的元工具

— 讨论

讨论.

评论区初始化中…

收束 · 一个人，一支 agent 队Closing · One Person, a Fleet of Agents.

回扣:六件与两翼The Recap: Six Pieces, Two Wings.

你只做机器替不了的两件事You Do Only What the Machine Can't.

边界,和一把不过期的尺The Boundary, and a Ruler That Doesn't Expire.

Codex · 多 Agent 的异步编码工作流Codex · The Async Multi-Agent Coding Workflow.

讨论Discussion.

收束 · 一个人，一支 agent 队.

回扣:六件与两翼.

你只做机器替不了的两件事.

边界,和一把不过期的尺.

Codex · 多 Agent 的异步编码工作流.

讨论.