Anthropic engineer shares Claude Code multi-agent orchestration playbook
English text is being prepared.
In a 30-minute workshop, an Anthropic engineer demonstrated an advanced use of Claude Code: running more than 100 agents simultaneously within one integrated loop, with 1 main agent in charge, while sub-agents act as project managers to break down tasks, distribute them, and cross-validate results. Based on this, the author suggests developers first use the main window to write task decomposition and acceptance rules, then spawn sub-agents to execute submodules concurrently. The author also warns that when a hundred agents run simultaneously, if termination conditions and permission boundaries are not hard-coded, it is easy to burn a large amount of tokens in an infinite loop, and beginners should start small with one main agent and three sub-agents.
Why read this
AI score 56
The original article recounts an Anthropic engineer's multi-agent orchestration approach and warns about token costs and permission-boundary risks when a hundred agents run simultaneously.
Topics
View original postThis is probably the fiercest yet most brilliant dimensional strike in the history of AI and the tech world. The day before, Anthropic was still loudly promoting that it would set aside 5% of its equity to establish a philanthropic fund to address the impact of technology, and Musk left a cold comment in the replies: I suggest you just give the shares directly to the public. Less than twelve hours later, he posted this tweet that stunned the entire internet: Next year, I will give every child in America one share of SpaceX stock I hold. This is just the first of many times. Based on the current market value, this is real money hitting more than 70 million families, with a total value exceeding 10 billion dollars. The most intriguing gap lies in these two completely different distribution paradigms. Traditional elite big tech in Silicon Valley loves to play with a refined set of moral rhetoric, putting shares into a charitable fund they control. On the surface it is giving back to society, but behind the scenes, how the money is spent, who it is invested in, and who draws the high salaries—all the power remains firmly in their own hands. Musk chose the path with the least room for retreat but also the purest one: truly cutting into his own personal assets, with no complicated intermediaries, directly distributing shareholder seats to every ordinary child at the bottom. This is why criticism can only ever bring arguments, but leading by example with action can end the game in an instant. When you accuse your opponent of empty posturing, the most forceful response is not to write a long essay to debate, but to pull out, within half a day, a set of real actions that the opponent simply dares not match and cannot afford to match, tearing away their moral fig leaf on the spot. What is even more formidable is the extremely far-reaching commercial gravitational field behind this move. More than 70 million children will hold a share of a rocket company flying into space in their accounts from childhood, which means that nearly every family in America will naturally become the company's most solid emotional community over the coming decades. Compared with spending billions of dollars on media PR and hollow advertising, this kind of move—binding interests directly to the fate of the next generation—is the truly insurmountable moat. People used to think that tech for good was empty talk written in social responsibility reports. From today on, someone has put the answer on the public screen: true goodness is not loftily setting up relief funds, but distributing the assets that belong to the future fairly, on day one, to the people who create the future. Do you think this approach of directly giving every child one share of hardcore assets, compared with the charitable funds run by a handful of elites at traditional big tech companies, is which one the future tech company should look like? Let me say this first: rather than watching institutions deduct large management fees in a black box every year, I would rather have a real rocket ownership share sitting under my child's name.
View original
这大概是整个人工智能与科技界历史上,最凶狠、但也最漂亮的一场降维打击。 前一天 Anthropic 还在高调宣传要拿出 5% 股权设立公益基金来应对技术冲击,马斯克在评论区冷冷留了一句:我建议你们直接把股份给公众。 结果不到十二个小时,他直接发了这条推文把全网看傻了: 明年把我持有的 SpaceX 股票,给全美每一个孩子送一股,这只是很多次里的第一次。 按现在的市值粗算,这是一笔砸向七千多万个家庭、总额超过一百亿美元的真金白银。 最耐人寻味的差距,就藏在这两种完全不同的分配范式里。 硅谷传统的精英大厂最喜欢玩弄一套精致的道德话术,把股份放进自己掌控的慈善基金,表面上是在回馈社会,背地里资金怎么花、投给谁、由谁领高薪,所有权力依然死死握在自己手里;而马斯克选择了一条最不留退路、但也最纯粹的路径:从自己的个人资产里真刀真枪地割肉,不设复杂的中间环节,直接把股东席位分给最底层的每一个普通孩子。 这就是为什么批评永远只能带来争吵,但行动示范能在瞬间终结比赛。 当你指责对手的假把式时,最暴力的回应不是写小作文去辩论,而是在半天之内掏出一套对手根本不敢跟、也跟不起的真实动作,当场把对方的道德遮羞布扯得精光。 更可怕的是这一步棋背后极度深远的商业引力场。 七千多万个孩子从小在账户里持有一股飞向太空的火箭股票,这意味着几乎全美所有的家庭在未来的几十年里,都会天然成为这家公司最坚固的情感共同体;比起花几十亿美元去买媒体公关和虚浮广告,这种把利益直接和下一代命运绑死的手笔,才是真正不可逾越的护城河。 以前大家以为科技向善是写在社会责任报告里的空话,从今天起有人把答案打在公屏上: 真正的向善不是高高在上地设立救济金,而是把属于未来的资产,在第一天就公平地分给创造未来的人。 大家觉得这种直接给每个孩子发一股硬核资产的做法,和传统大厂搞的那些由少数精英掌管的慈善基金相比,哪一种才是未来科技公司该有的样子?我先说,与其看机构在暗箱里每年扣掉大笔管理费,我宁愿自己孩子名下躺着一张真实的火箭所有权。
Body
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。
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