An Anthropic engineer shares Claude Code multi-agent orchestration approach

If you're still using Claude Code by asking one question and having it change one line, you're basically just treating it as an advanced version of Google Search. An Anthropic engineer just revealed an advanced internal approach in a 30-minute workshop: running more than 100 agents simultaneously in one integrated loop, with one main agent in charge and all the sub-agents acting as project managers leading teams. Those so-called Vibe Coding training courses selling for $500 on the market combined aren't worth as much as this dense 30-minute workshop just released by an Anthropic engineer, who explained in detail how to squeeze Claude Code dry at the lowest operating cost. In the sharing, this internal engineer showed off his extremely hardcore daily setup: he runs more than 100 agents simultaneously in one integrated loop, with only one main agent in charge at the top, while the sub-agents below each act as project managers, leading and coordinating the entire R&D team themselves. Normally, when we open Claude Code to write code, the default action is one question and one answer; if there's an error, we paste it in and ask how to fix it. In essence, we're still using it as a Google Search that can write code, with one person watching one terminal window—when you stop, it stops too. But this Anthropic engineer pointed out the real dividing line that creates a tenfold productivity gap. Claude Code's real nuclear weapon has never been single-window conversational Q&A, but running it as a self-learning multi-agent swarm. While you're still hand-holding an intern to correct typos, people in the know have already spun up an entire 24/7 virtual engineering team of a hundred people in the background. It's like building a house on a construction site: the ordinary way is you swing the hammer yourself and have the AI hand you a nail for each strike, whereas his main-agent-plus-project-manager architecture directly upgrades you to general contractor. You only need to align on the blueprints with the chief foreman, while the project managers below split modules, distribute tasks, and cross-check acceptance themselves, putting the entire project into an automated closed-loop pipeline. For everyone who heavily uses Claude Code and Cursor to write code every day, this is a level-up move you can copy right away. Starting tomorrow, stop piling all the context into one long conversation window until it explodes. Try first writing the task breakdown and acceptance rules in the main window, then letting it spawn sub-agents to execute different submodules concurrently, completely switching your role from a coder who personally types code to an architect who only looks at the final test results. However, while enjoying the thrilling picture of a hundred agents running together, you also need to stay clear-headed. Although he emphasized in the video how to keep operating costs down, in a real project, once you start an integrated loop of a hundred agents, if the main agent's termination conditions and permission boundaries are not hard-coded, it is extremely easy to burn through your token bill in an infinite loop. Moreover, as an Anthropic engineer himself, his latitude in allocating internal compute is far beyond that of ordinary pay-as-you-go developers. For beginners, the safest path is to start by practicing with a small swarm of one main agent and three sub-agents. Treating a top-tier coding agent as a search engine for Q&A is like buying an entire automated fleet but only taking the flagship out to row on a park pond every day. The core multi-agent architecture and cost-reduction points from this 30-minute workshop are compiled in the first floor [comment]; I suggest watching it together with the original video. https://t.co/OAp6oUuseD

View original

如果你用 Claude Code 还是你问一句、它改一行，那基本只是把它当成了高级版的 Google 搜索引擎， 一位 Anthropic 工程师刚在 30 分钟的工作坊里亮出了内部高阶打法：在一个集成循环里同时跑 100 多个智能体，1 个主代理坐镇，底下全当项目经理去带团队。 市面上那些卖 500 美元的所谓 Vibe Coding 培训课，含金量加起来都不如 Anthropic 工程师刚放出的这场 30 分钟高密度工作坊，他直接把怎么用最低的运营成本榨干 Claude Code 给讲透了。 这位内部工程师在分享里晒出了自己极其硬核的日常配置，他在一个集成循环里同时跑着超过 100 个智能体，上面只设一个主代理坐镇，底下的子代理分别充当项目经理，自己去领导和调度整个研发团队。 平时我们打开 Claude Code 写代码，默认动作是一问一答，报错了就贴进去问怎么改，本质上还是把它当成一个会写代码的 Google 搜索引擎在用，一个人盯一个终端窗口，你停下来它就跟着停下来。 但这位 Anthropic 工程师点破了真正拉开十倍产能差距的分水岭，Claude Code 真正的核武器从来不是单窗口的对话问答，而是把它当成一套能自主学习的多智能体蜂群去跑，当你还在手把手给一个实习生改错字时，懂行的人已经在后台拉起了一整支全天候运转的百人虚拟工程队。 这就像你在工地上盖楼，普通用法是你亲自抡着锤子，让 AI 递一颗钉子你敲一下， 而他的主代理加项目经理架构，是直接把你升格成总包商，你只管跟总工长对齐图纸，底下的项目经理自己去拆分模块、分发任务、交叉验收，把整个工程装进一个自动闭环的流水线里。 这对所有每天重度用 Claude Code 和 Cursor 写代码的人是一个马上就能照抄的升维动作，明天起别再在一个长对话窗口里把所有上下文堆到爆了， 试着先用主窗口写好任务拆解与验收规则，再让它派生子代理去并发执行不同的子模块，把你自己的角色从亲自敲代码的码农，彻底切换成只看最终测试结果的架构师。 不过看着百个智能体同跑的爽文画面也得留个清醒，他在视频里虽然强调了怎么压低运营成本，但在真实项目里一旦开启百个代理的集成循环，如果主代理的终止条件和权限边界没写死，极容易在死循环里把 Token 账单烧穿， 况且他作为 Anthropic 自家工程师，调配内部算力的宽裕度远非普通按量付费的开发者可比，新手上来先从一主三副的小蜂群练手才是最稳妥的路子。 把顶级编程智能体当成搜索引擎一问一答，就像买了一整支自动化舰队，却每天只拿旗舰去公园划船。 这场 30 分钟工作坊的核心多代理架构和压低成本要点我整理在一楼了，建议配合原视频一起看。 https://t.co/OAp6oUuseD