为什么一个 AI 智能体不够：Anthropic 工程师详解千智能体编排

简要说明Brief Description

为什么一个 AI 智能体还不够（以及如何运行 1,000 个）——CJ Avilla 和 Harrison Stall，AnthropicWhy One AI Agent Isn’t Enough (And How to Run 1，000) - CJ Avilla & Harrison Stall， Anthropic

目录Table of Contents

• Unix 与 OS/360：扩展性问题
  Unix vs. OS/360: the scaling problem
• 认识 CJ 和 Harrison
  Meet CJ & Harrison
• Claude 托管智能体中的流程介绍
  Introducing Flows in Claude Managed Agents
• 为什么单个智能体会遇到瓶颈
  Why single agents hit a wall
• 动态工作流的工作原理
  How dynamic workflows work
• 演示：股票观察名单扫描器
  Demo: stock watchlist scanner
• 从控制台到生产环境：调整各个阶段的配置
  From console to production: tuning each phase
• 问答，第1轮
  Q&A, round 1
• 漏洞猎手：多智能体 PR 审查
  Bug Hunter: multi-agent PR review
• Explainer Studio：制作视频的流程
  Explainer Studio: flows that make videos
• 惰性、自我偏向与目标偏移
  Laziness, self-bias & goal drift
• 大规模代理安全性
  Agent safety at scale
• 声明、包含、检测
  Declare, contain, detect
• 将大脑与双手分离
  Separating the brain from the hands
• 凭据保险库与工具锁定
  Credential vaults & locking down tools
• 总结与问答，第二轮
  Wrap-up + Q&A, round 2

Unix 与 OS/360：扩展性问题Unix vs. OS/360: the scaling problem

主持人：1969年，离这里大约30英里的地方，Ken Thompson当时坐在贝尔实验室里。他的妻子刚带着他们一岁的儿子去加利福尼亚看望他的父母，要待一个月。于是他坐下来，说：“我有一个月时间，还有一个想做的项目。我要把它拆分开来：用一周构建内核，用一周构建编辑器，用一周构建汇编器。然后呢，再用一周构建另一项东西。”一个月结束时，他就有了Unix。明白吗？Ken Thompson经历的这个过程有一个很妙的地方，那就是他拥有所谓的“概念完整性”。也就是说，一个人能够把整个项目都装在脑子里，并从头到尾独立完成整个项目。Host: In 1969, just about 30 miles from here, Ken Thompson uh was sitting at Bell Labs and his wife had just taken their 1-year-old uh son out to California to visit his parents for a month. And so, he sat down and he said, "I have a month and I have this project that I want to do. I'm going to carve it up and in 1 week, I'm going to build a kernel. In 1 week, I'm going to build an editor. In 1 week, I'm going to build an assembler. And then, uh you know, 1 week, I'm going to build uh uh one other thing." And he had at the end of that month, he had uh Unix. Okay? So, what what's beautiful about this process that Ken Thompson went through is that he had what's called conceptual integrity. And that means that one person can hold the entire project in their head and execute the entire project end to end.

还有另一个操作系统项目，嗯，也是一个规模巨大的工程，发生在 20 世纪 60 年代和 70 年代初期，当时 IBM 正在尝试开发 System/360 和 OS/360。此前，如果你想升级 IBM 的计算机，每次升级硬件，都得把随之配套的所有软件重新编写一遍。因此，OS/360 是一项规模庞大的工程，他们确实是押上了整个公司的未来，还雇用了大批工程师。他们坐下来，做了些粗略的估算，说：“好吧，这大概要耗费 5,000 人年。那我们就雇人，派多达 1,000 名工程师来做这个项目，看看能不能在很短的时间内完成。”这个项目也带来了很多经验教训。There's another operating system uh that uh was uh you know, a huge effort and that happened also in the '60s and the early '70s at IBM when they were trying to build System/360 and OS/360. And this was an undertaking where previously, IBM computers, if you wanted to upgrade your computer, every time you upgraded the actual hardware, you had to rewrite all of the software that came along with it. And so, OS/360 was this massive effort where they really bet the company and they hired tons and tons of engineers and they they sat down just some back of the map back back of napkin math and said, "Okay, this is going to take about 5,000 sort of man-years. So, let's hire and throw as many as 1,000 engineers at this project and see if we can get it done in uh in a a short window." And there was a lot of learnings that came out of that project.

Fred Brooks 是这个项目的经理之一。项目结束后，他说：“好，我们学到了很多东西。把它们记下来吧。”于是，他把这些经验写进了一本名为《人月神话》的书里。这本书非常受欢迎，讲的是如何组织和管理团队，才能构建并交付 OS/360 这样规模庞大的项目。其中一个重要结论是，组织团队的一种非常有效的方式，有点像外科手术团队：由一位领导者或首席专家带领，周围有 10 位专家。比如，团队里可以有外科医生，以及语言专家、助手、副驾驶员等；这些人各有所长，但外科医生最终要把所有事情都了然于胸，从而维护整体的概念完整性。Fred Brooks was one of the the managers of the project and he came away and said like, "Okay, we learned a whole bunch of stuff. Let's write it down." So, he wrote it in in this book called The Mythical Man-Month. So, it's like a very popular book. And this is a bunch of organizational learnings and management learnings from how to build and ship this really massive project in OS/360. One of the big takeaways was that a really effective way to organize a team is kind of like a surgical team where you have one sort of leader or one chief, and they're surrounded by 10 experts. And so, you might have you know, the surgeon, and then they have you know, a language lawyer, and they have a sidekick, and they have a co-pilot, and then all of those people have their specialties, but the surgeon ultimately has everything in their head, and they can maintain that conceptual integrity.

这很棒，你们用这个小团队就能创造奇迹。外科医生一直都在这么做。但我们也想采用同样的结构，并将其大幅扩大。快进到大约一个月前，也就是几个月前，我们看到 Linux 7.2 发布了。这是另一个操作系统项目。这个项目历时九周，我们收到了来自 300 家不同公司的 2,600 多位开发者的贡献。而一开始并不是这样的。也就是说，Linux 的实际贡献方式一开始并不是这样，但我们逐渐演变出这种结构，让大家能够以这种方式为 Linux 做贡献。所以，如果你回头看看 1998 年的内核邮件列表，就会发现 Linus 当时好像有点崩溃。他收到别人发来的补丁，他们会给他发邮件，然后你知道，还会说：“我已经给你发了三次补丁了。And that's great, and you can perform miracles. Surgeons do it all the time with this small team. But we also want to take that same sort of structure and scale it up much further. We fast-forward to just a a month ago, a couple months ago, we saw Linux 7.2 was released. So, this is another operating system project. This took nine weeks, and we had contributions from 2,600 something different developers across 300 different companies. And it didn't start out that way. Like, the actual contribution to Linux didn't start out that way, but we sort of evolved the way that we can contribute to Linux into this structure. And so, if you look back at the kernel mailing list from 1998, you'll see that Linus was actually like kind of having a bit of a breakdown. He was receiving patches from people they were kind of like emailing him, and then you know, saying like, "I've emailed you three times with my patch.

你为什么还没把这个提交到那个——你知道的，内核里，或者别的什么地方？”他就说：“够了。我要去度假了。我不想再听这些。别再往我的收件箱里发邮件了。”你知道，他差不多到了忍耐的极限。我觉得，现在我们当中很多人都在协调和管理大量智能体，可能也会开始有这种感觉：自己成了瓶颈，快要到崩溃的边缘了。因为有大批智能体提交了成堆的 PR，而我们就坐在那里，在 GitHub 上试着审核 PR，确保一切都按预期运行。呃，直到几年后，我们才真正做出了改变。不过，就在 Linus 发出这条大意是“我受够了，我要去度假”的消息的第二天，另一个人——Larry McVoy——发帖说：“问题在于，Linus 无法扩展。Why haven't you submitted this to the you know, to the kernel or whatever?" And he's like, "That's it. I'm going on vacation. I don't want to hear it. Get out of my inbox." And sort of you know, he he he hit his breaking point. And I think a lot of us who are orchestrating and managing lots of agents now might start to feel this feeling of the being the bottleneck and being the breaking point as we have lots of agents contributing tons and tons of PRs and we're sitting there, you know, in our in our GitHub trying to like approve PRs and make sure that everything works as expected. And so uh it wasn't until a couple years later that we actually changed. But actually the day after Linus posts this like kind of uh you know, I'm done, I'm going on vacation, uh another person posts and they Larry McVoy and he says, "The the problem is that Linus doesn't scale."

Larry 也稍微自我推销了一下，他说：“我有个主意。我们用版本控制怎么样？”他当时有点像是在推销这个商业产品。大概直到四年后，也就是 2002 年左右，Linus 才说：“好吧，你知道吗？确实如此。我没法扩展。与其采用星形结构，让每个人都直接向我贡献代码，我们更需要以网络的方式协作，建立起层层信任关系，让大家把贡献提交给值得信赖的维护者，由他们来审查代码。然后，只有那些特殊情况才会被逐级上报。不过，Linus 最终仍然可以是那个向 main 提交代码的人。”左边可以看到，这种星形结构中有大约 36 个参与者，都在以某种方式贡献代码。And uh Larry also did a little bit of self-promotion and he's like, "I have an idea. How about we use version control?" And was kind of like pushing this commercial product. Uh and so it actually wasn't until maybe four years later around 2002 that Linus was like, "Okay, you know what? It's true. I don't scale. And instead of uh operating as a star structure where everyone is sort of contributing directly to me, we need to work more as a network where we have these layers of trust and people can submit contributions to trusted maintainers who are going to review that code. And then only those exceptional cases will be surfaced up. Uh but Linus can still kind of like ultimately be the person that is uh committing to main." On the left you'll see this star structure has like 36 different actors that are all kind of contributing in.

右边最终形成了这样一种挺不错、挺漂亮的结构，每一层往上汇报这些变化的人大概最多只有四个。呃，在旧金山，如果你参加了 AI Engineer 大会，就听过 Tarek 和 Gary 的分享，他们传达的信息是：“我们有 Fable。这真的令人兴奋，现在是时候不讲常规了，是时候更有雄心壮志，去打造更大的东西了。”Gary 还告诉大家，现在是时候把海洋煮沸了。因此，我们拥有这些能力非常强大的模型，今天也很高兴能聊聊，我们如何把这些模型编排成能够彼此协作的系统，从而构建真正、真正庞大的项目，解决更大的问题。On the right we end up in this like nice beautiful structure that only has like maybe at most four people at each level sort of reporting those changes up. Uh in San Francisco, uh if you were at AI Engineer you heard from Tarek and Gary and the message was, "We have Fable. It's really exciting and it's time to be unreasonable. It's time to be more ambitious and build bigger things." And Gary told you it's time to boil the ocean. And so we have these very very capable models and we're excited to talk a little bit today about how we can orchestrate these models into uh systems that can collaborate with each other to build really, really massive uh projects and and solve even bigger problems.

认识 CJ 和 HarrisonMeet CJ & Harrison

那么，我是 CJ Avilla，是 Anthropic 的技术团队成员，而这是……So, I'm CJ Avilla. I am a member of technical staff at Anthropic and this is

嘉宾：我是 Harrison，也是 Anthropic 的技术团队成员。大家早上好！感觉怎么样？兴奋吗？不错，气氛很热烈。太有劲儿了。哇哦！好，那我们开始吧。我先简单介绍一下本次环节的安排。我们会先花些时间告诉大家 Claude 托管智能体到底是什么、我们要聊些什么，以及这个多智能体编排引擎究竟是怎么回事。我们会花一会儿时间来讲解它。环节中也会有动手实践的工作坊，所以请拿出笔记本电脑。稍后幻灯片上会介绍开发者额度，这样大家就能亲自上手试试这些我们非常兴奋、即将推出的功能。不过在此之前，我们先稍微概述一下这里到底要做些什么。Guest: Harrison, and I'm also a member of technical staff at Anthropic. And good morning, everybody. How we doing? Excited? Nice. Lot of energy. Lot of energy. WOO. UH ALL RIGHT. WELL, AS WE GET GOING HERE, um just want to give you guys a quick lay of the land of how this session's going to run. Um it's going to be some part like telling you what the heck even is Claude managed agents and what are we about to talk about and like this whole multi-agent uh orchestration engine that we're uh going to spend a little while explaining. Um it's also going to be part hands-on workshop, so um laptops out. We'll have uh developer credits uh on on slide later on, so you can have actually get your hands dirty. Uh try now some of these features that we're super excited to ship. Um but before that, we uh we uh we'll give you a a little bit of over of an overview on what the heck is going on here.

所以，嗯，我们非常期待周五能……So, on Friday, um we are so excited to

Claude 托管代理中的 Flows 简介Introducing Flows in Claude Managed Agents

在 Anthropic 大约四月份推出的一个名为 Claude managed agents 的产品界面中，我们宣布支持一项名为 flows 的新功能。Flows 本质上就像是多智能体编排的下一代版本，你可以让 Claude 全面编排自己的整个多智能体拓扑结构。就像 CJ 刚才提到的那些非常复杂、相互关联的网络，它们映射了现实世界中的组织结构。我们设想，未来 Claude 基本上能够自行编排，对吧？而 flows 就是我们在 API 产品中实现这一点的方式。所以，在深入探讨 API 产品和具体体验之前，你们中很多人可能都熟悉 Claude Code 中的动态工作流。你知道，你输入：“嘿，Claude，去做 X、Y 和 Z。”announce within a product surface called Claude managed agents that Anthropic ships or shipped back in um April that we support a new feature called flows. Flows are essentially like the next version of multi-agent orchestration, where you can let Claude fully orchestrate the entirety of its own multi-agent topology. So, like CJ was just talking about these really complicated interconnected webs uh that map to organizations in the real world, we envision that there's a future where Claude is able to basically orchestrate itself, right? And uh flows are the way that we do that on our API surfaces. So, before we get too deep into uh the API surface and the experience specifically, um a lot of you are probably familiar with dynamic workflows in Claude code. Um, you know, you type in, "Hey Claude, like go and do X, Y, and Z.

“使用 ultra code。”它会启动一个相当复杂的智能体网络，并由它自行完成全面编排。它底层使用了一个非常有意思的 JavaScript 引擎，基本上可以在 Claude Code 深处支持类似 map-reduce 的任务。嗯，我们知道，这已经成了一个非常非常受欢迎的产品，人们可以在终端里输入指令并实时使用它。嗯，幻灯片上有一些例子，但关键在于，你不必真的把这种分类操作，或是分发任务再综合结果、进行对抗式验证的拓扑结构画出来——比如说“这是智能体一，它负责做这些；这是智能体二；然后等智能体一采取某个行动，诸如此类”。嗯，你完全可以直接告诉 Claude 使用 ultra code 来达成目标，对吧？Use ultra code." It spins up this like pretty complicated web of agents. It's fully orchestrated itself. It uses a really interesting JavaScript engine under the hood to basically support like map reduced ask like activities in deep within Claude code. Um, and we know that this has become a really, really popular product uh for like typing into a terminal and like using it live. Um, so there's a bunch of examples on slide, but um, the idea is that instead of having to like literally draw out this classifying act or fan out and synthesize adversarial verification style um, like topology where you say like, "This is agent one and it does this and this is agent two and like wait until agent one takes some action and yada yada." Um, you can literally just tell Claude to use ultra code to uh achieve its goal, right?

Claude 会根据我们在内部进行的模型训练，以及它自身的能力，自然而然地知道如何有效地协调自己，从而完成越来越困难的工作。嗯，所以我们觉得，这让我们初步尝到了未来的滋味：我们将如何利用完全由智能体驱动的自动化，解决一些我们认为是人类最棘手的问题。这个嘛，我不知道，我确实特别喜欢琢磨这些东西，但我对这个方向的发展真的、真的非常兴奋。嗯，动态工作流以及更广泛来说，这种……真正酷的地方是艰。And Claude will just know based on uh model training that we've done internally and generally just uh its own capabilities, how to effectively orchestrate itself in order to accomplish harder and harder work. Um, so we we think this is like we're getting a taste of how we're going to solve some of what we think are humanity's hardest problems using like fully agentic automation. And so this is like, I don't know, I I definitely nerd out about this stuff, but I'm like so, so excited about the direction that this is even headed. Um, what's really cool about dynamic workflows and just generally like this

为什么单个智能体会遇到瓶颈Why single agents hit a wall

你使用多智能体拓扑，而不是只用单个智能体，是因为单个智能体在面对难度低得多、野心小得多的任务时，就会早早撑不住。嗯，所以这些数字只是粗略估算，对吧？单个智能体可以运行一个小时。协调器加智能体名单，是我们对多智能体的第一版尝试：有一个单独的协调器，下面只有一层智能体，它们可以各自去完成非常具体的任务。这样可以运行得久一点。但我们认为，动态工作流才是解决真正困难问题的关键，它支持任意复杂的嵌套智能体集合，让这些智能体并行运行、彼此交互。所以，我们花了很多时间和用户交流。我也特别期待今天能和大家聊聊，因为我想你们当中有些人对行业最新动态也有不少独到见解。you using multi uh multi-agent topologies rather than just a single agent is that it a single agent keels over much, much earlier with much less ambitious tasks. Um, so uh you know, hand waved numbers here, right? Uh a single agent runs for an hour. A coordinator and roster, which is like our first pass at multi-agent where we had like kind of this like single coordinator that like has just one level of agents underneath it that can go off and accomplish very specific tasks um, can run a little bit longer, but then dynamic workflows, we think this is like the key to solving actually hard problems, which enables arbitrarily complex nested sets of uh of agents operating all in parallel and interacting with each other. So, we spent a lot of time talking to our users. Um and I'm also like super excited to talk to all of you guys today because I imagine some of y'all have some hot takes on the latest in the industry as well.

而且我们一直在听到很多相似的主题。嗯，首先，你们正在做一些很厉害的事情，构建自己的多智能体运行框架，对吧？你们正在搭建极其复杂的系统，来解决我们刚才谈到的单智能体面临的一些问题。嗯，你们在手动编排多智能体，还得处理容器重启，对吧？比如说，如果一百个智能体中的一个在任务进行两小时后挂掉了，你们要怎么从它中断的地方继续？又要如何确保大致了解整个拓扑的状态，等等？嗯，做到这些往往并不容易。另外，嗯，幻灯片左下角这一点大致是说，单个智能体可能得追踪它下面的 60 个智能体，就拿这个数字举例来说。And we we've we've been hearing a lot of common common themes throughout. Um so, first of all, you guys are doing some crazy stuff building out your own multi-agent harnesses, right? You've got incredibly complicated systems that you're building in order to basically solve for some of those problems that we were talking about just a second ago with the single agent. Um so, you're you're hand rolling multi-agent orchestration. You're having to handle container restarts, right? Like what if one of your agents out of a hundred dies uh you know, like two hours into the task. How do you actually pick back up where that agent left off? Make sure you understand generally the state of the whole topology, etc. Um and it's often like very non-trivial to do so. Um additionally, we uh this this bottom left point in the slide is basically like you're having a single agent have to keep track of like 60 as an as an example agents underneath it.

就像 CD 刚才谈到的，Linus 这种方式并不能无限扩展。你知道，他是一位非常出色的协调者，但遗憾的是，单个智能体能记在脑子里的东西总是有限的。最后还有一点：你得想办法检查自己构建出来的智能体图。你必须逐个去查看图里的节点，问它们流程进行到哪一步了吗？还是把所有事情都交给协调者？一般来说，你要怎么随时了解整个图的状态呢？ 我们已经听过使用 Cloud Code 的人提到这个问题了，对吧？我们的 API 最初几个版本也是如此。我记得有个梗，大概是在 Opus 4.5 发布之后吧，模型的能力突飞猛进，让自主编码变得极其强大，于是人们走来走去时都会把笔记本电脑稍微打开一点，免得 Cloud Code 退出，还要确认智能体仍在运行。And like CD was talking about Linus doesn't scale like that, you know, he's an incredible coordinator, but unfortunately, there's just like some cap to how much the single agent um is able to keep in their head. And then last but not least, you have to figure out how do you even inspect the graph of agents that you've you've built out. Do you have to go and like poke at each of the nodes in the graph and like ask where it's at in the in the process? Do you run everything through the coordinator? Generally speaking, like how how do you understand the state of the graph at any point in the process? Um so, we've already heard this just from people using Cloud Code, right? Um and our first renditions of our own APIs as well. Don't let me remember there there was this meme that when I think it was probably like post Opus 4 5 when when the models really took off and the the capability just made autonomous coding so so so powerful where people were like walking around with their laptops slightly open so that their cloud code wouldn't die and like making sure that their agents were still running.

这就像让人在流程中不停地点“批准、批准、批准、批准、批准”，只是为了让事情继续运转。我们认为这有很大的局限性，对吧？如果我们想解决真正棘手的、关乎整个社会的问题，这种做法根本无法随着我们的雄心壮志而扩展。接下来，我们来看看 Claude 托管代理，以及我们如何在自己的生态系统中嵌入这种复杂的多智能体拓扑结构。我想，在座的很多人可能……不如我直接问吧：有多少人听说过 Claude 托管代理？It's like people in the park like pressing two like approve approve approve approve approve just to keep the the thing running. And we think that that's like hugely limiting, right? Like that is if we want to solve truly hard like society level problems, that's just not going to scale for our ambitions. So fast forward to Claude managed agents and how we've embedded this kind of complex multi-agent topology within our own ecosystem. So I recognize that a lot of you in this room probably actually let me ask how many of you have heard of Claude managed agents?

主持人：[嗤笑声]Host: [snorts]

嘉宾：哇哦，这其实相当不错。大概有半个房间的人都在用。你们当中还没用过的人，应该去看看。我觉得过去一年左右，它基本上成了我每天脑子里想的全部内容。简单来说，它是一套 API，让 Anthropic 来管理智能体的生命周期。你可以定义一个智能体，基本上就是一组技能、工具、MCP 服务器等等。然后在一个会话中实例化这个智能体——会话本质上就是一个用来完成实际工作的请求。接着，智能体就会自行运转，按照它要完成的目标大展身手。我们提供了一整套 API，让你能够存储这些有状态的 API 资源，并通过持久化事件日志实时监控一切。所以，去看看 Claude 托管式智能体吧。我们今天当然就是来和大家聊聊这个的。Guest: Oh wow, that's actually pretty solid. It's about 50% of the room. For those of you who are not in fact did you should check it out. It's also like all I think about every day of my life basically for the past year or so. But it's essentially an API surface that lets you have Anthropic manage the the agent life cycle. So you define an agent it's basically like a list of skills, tools, MCP servers, etc. You instantiate that agent within a session which is like a request essentially to to get real work done. And then that agent just goes off and goes off and cooks based on whatever it is that it's trying to accomplish. So we have a whole API surface that makes it possible to store these like stateful API resources and then monitor everything in real time via durable event log. So yeah, check it out Claude managed agents and what we're here of course to to talk you talk to you all about it.

嗯，不过回想起来，我觉得是在五月，我们发布了所谓的多智能体 V1，对吧？在这种模式下，你必须定义主代理能够启动的每个子代理。然后，你基本上还得监听我们所说的每个代理正在运行的线程，对吧？所以，嗯，这很像操作系统，对吧？整个进程有不同的线程，每个线程都有自己的上下文。嗯，它们之间会共享一些状态，但通常要共享状态，你就得管理所有线程之间来回发送的消息。嗯，我想说的是，尽管我们的 API 接口支持这种模式，但配置起来并不是件特别简单的事。Um but back in I think it was May we shipped what we called like multi-agent V1, right? And this is where you have to define each of the sub agents that your primary agent is able to spin up. And then you have to listen to basically all of what we call the threads that each of those agents are operating on, right? So, um, it's a lot like, you know, an operating system, right? Where you have the the overall process has different threads, each of those threads have their own context. Um, there is some uh state that's shared between them, but typically in order to share state, you have to like manage message message sends uh back and forth between all the threads. Um, my point in saying all this is that um it it wasn't the most uh trivial thing to set up even though our API surface supports it.

嗯，不过快进到上周五，我们发布了这个东西。Um, but fast forward to last Friday, um, where we shipped this

动态工作流的工作原理How dynamic workflows work

云托管代理中有一种叫作动态工作流的新功能，我们认为，通过这种非常简单的方式，可以更轻松、更有效地发挥这些能力日益增强的模型的潜力。它的工作方式是：你会有一个主代理，负责制定计划，决定如何完成任务。比如说，把我的百万行代码单体仓库从 Python 迁移到 Rust。嗯，这个代理接着就能启动……一旦它启动并制定了计划，就可以根据自身的判断能力或能力水平，自动调度一组由它创建、并可按需启动或关闭的代理。在多代理编排的多个不同阶段中，它能够完成一系列离散且可验证的工作单元，而你不必管理这些代理之间的任何连接。new thing called dynamic workflows within cloud managed agents, and we think that this is like way easier and way more possible to get power out of the these increasingly capable models in a very trivial way. So, the way that it works is that you have a lead agent that is essentially responsible for creating a plan for how it wants to accomplish a task. So, you know, migrate my million line mono repo from Python to Rust as as an example. Um, that agent is able to then spin up Once that agent spins up it its plan, it's able to automatically fan out to its own set of agents that it spins up and down based on uh its own uh uh I I guess rationality or um capability, and across multiple different phases of the multi-agent orchestration, it's able to accomplish like discrete, verifiable units of work without you having to like manage any of the interconnections between these things.

嗯，基本上来说，可以把它分成几个阶段。第一阶段，比如说读取文件，然后启动一批探索代理，让它们去处理。第二阶段，也许是——我也不确定——找出这些文件中的某些信息。显然，如今这些代理比我聪明得多，会想出比我临场能想到的更好的方案。不过最后，我们认为，这基本上就是通过 API 接口，在平台上完成非常非常非常艰巨工作的途径。所以，如果你想把这样的功能集成到自己的产品中，这是一种非常简单的实现方式，不用想办法再次打开 Cloud Code，或者把它连接到你自己的系统。而且，在这些动态工作流运行时，你首先可以非常轻松地实时观察正在发生的事情。Um, so, very basically, it might be like phase one, like read the files, and it spins up a bunch of explorer agents to go off and do so. Phase two, maybe I I don't know, identify something about those files. Obviously, the the agents are way smarter than me uh these days and would would come up with a better plan than than I could on the spot. Um, but then at the end, um, you we think that this is essentially like the path to getting really, really, really hard work done on platform via an API surface. So, if you want to build like this into your products, this is a really trivial way for you to do so without you having to somehow like have Cloud Code open again and I don't know, like connect that to your to your own systems. So, while these dynamic workflows are running, it's very trivial to first of all watch what's happening in real time.

有一个统一的事件流，基本上可以实时展示各个智能体正在做的一切。你可以逐个查看智能体，了解实时发生了哪些事件。除此之外，我们知道 token 消耗过多确实是个问题，所以你可以在启动会话之前为其设置预算，这样就不会消耗大量 token。再比如，如果你有某个系统在监听智能体实时执行的操作，而你觉得它们的发展方向不太理想，或者想实时引导它们，就可以通过我们的 API 接口中断并停止它们。相比之下，Cloud Code 可能会让人有点困惑，因为它们都是工作流，对吧？There's a single event stream that basically gives you everything that the agents are doing in real time. You can check agent by agent what the events what events are happening in real time. On top of that, we know that token hunger is like a real problem and you can set budgets on your sessions before kicking them off so you don't actually only burn substantial amounts of tokens. And then if you let's say that you have like some system that's listening to generally what the agents are doing in real time. You feel not so excited about the direction that they might be going or you you want to steer them in real time, then you can do so by just like stopping them by interrupting via our API surface. So, in comparison to Cloud Code, I I recognize that like this can be a little bit confusing because the they're both workflows, right?

不过，我们今天讨论的是 Cloud Managed Agents 中的这些工作流。我们提供的是一组 API 接口，你可以将它集成到自己的应用界面、内部流程，或者你想把 Cloud 嵌入的任何系统中。你不必非得在终端里输入命令，或实时与它交互。也许你想把它接入 CI，对吧？这就是最简单的做法。 在这里的对比表中，我想特别指出的是，我们看到很多用户会按计划运行这些任务，或者监听 Webhook。GitHub 就是一个很典型的例子：你可能想构建一个自动搜寻 Bug 的工作流，对吧？稍后我们会实际聊聊，也会演示一下。But we're talking today about these workflows within Cloud Managed Agents. So, we're an API surface that you should be able to integrate into your own application surfaces or your own internal processes or your own, you know, whatever it is that you might want to wherever you want to like embed Cloud within your systems. You don't have to be like typing on a terminal or interacting with the thing in real time. Maybe you want to hook it up to CI, right? Like this is like the the trivial way to do so. Um So, in our comparison chart here, I think the thing that I would highlight is that we we see a lot of users using like running these things on schedules, listening to webhooks. GitHub example is a pretty pretty classic one of like maybe you want to build out a bug hunter workflow, right? We'll actually talk about and and show show off a little bit later.

嗯，是的，我想这大概就是我想在这里分享的内容。那么，话不多说，我们就开始做其中一件事吧。嗯，Um and yeah, I think that that's probably about what I want to share here. Um, so with that said, let's go ahead and start one of these things. Um,

演示：股票观察列表扫描器Demo: stock watchlist scanner

我说够多了。来，呃，看看吧。是那个吗？对。对，过渡很流畅。enough of me yapping. Let's uh, let's check it out. Is that Yeah. Yeah, seamless transition.

主持人：我看看。你能看到那个吗？嗯，那里……Host: Let's see. Can you see that? Yeah, there

嘉宾：好，太棒了。大家都准备好了吗？好，我把画面稍微放大一点。嗯，我也知道，在座有些人可能之前从没见过 Cloud Console，所以我先花一分钟简单介绍一下。大家稍微耐心听我说一下。对于 API 集成者来说，Cloud Console 是管理 Anthropic API 服务中几乎所有资源的地方。在控制台里，我现在位于默认工作区，可以看到这个不错的下拉菜单。我稍微缩小一下画面，这样你们就能看到左侧这个不太好看的下拉菜单，叫作“管理代理”。“管理代理”产品的所有 API 资源都在这里。我们今天也会主要在这里操作。嗯，点击“快速入门”后，你会看到一大堆基本上只需点一下就能上手的工作流，可以直接运行。Guest: All right, sweet. All right, we all good? Okay, I'm going to make it a little bit bigger. Um, so I also recognize that some of y'all might not have ever seen the Cloud Console before, so just I'll spend a minute here talking a little bit about it. So so so so hang with me here. The Cloud Console is where you as an API integrator can manage essentially all of your resources within Anthropic's API service. So in the console, I'm in the default workspace and I'll see this nice drop-down. I'll zoom out a second just so you can see on the left side this not nice drop-down called manage agents. This is where all of the API resources live for the manage agents product. It's also where we're going to be spending our time today. Um, if you click into the quick start, you can see that there's a whole bunch of basically one click to get you going kinds of uh, workflows that you can run.

不过，如果你滚动到最底部，就会看到这里有一个挺不错的小标签，叫作“动态工作流”。这就是我过去几分钟一直在讲的内容。考虑到我知道在座有不少金融领域的朋友，我们来运行一下股票观察列表扫描器动态工作流。嗯，所以你可以看到，这个配置的作用基本上是创建一个使用 Claude Opus 5.5 的代理，传入一种名为 multi-agent 2026-10-01 的类型，然后提供一个系统提示词，大致说明代理本身是什么，也就是代理的角色设定。嗯，我就不在这上面花太多时间了，不过，简单来说，为了让大家了解一下，我们点击“使用此模板”时，它会在底层创建代理资源。那我们开始运行吧，看看效果如何。But if you scroll to the bottom, you'll see this nice little tab here called dynamic workflow. And that's what I've spent the last few minutes talking about. So because I recognize there is a a bunch of finance oriented folks in the room, we're going to run the stock watch list scanner dynamic workflow. Um, so you [clears throat] can see what this config is doing is essentially creating a agent that's using Claude Opus 5.5 passing in this type called multi-agent 2026-10-01 and then providing a system prompt that is basically like what what the agent itself is is uh, what the persona of the agent is. Um, I'm not going to belabor this too much, but um, basically the what for all of your edification, this is creating the agent resource under the hood when we click use this template. So, let's go ahead and get it rolling and see see how this goes.

嗯，首先，您的智能体要想在现实世界中完成任何事情，都需要一个环境，对吧？本质上，这是一台供 Claude 操作的计算机。嗯，过去大概 10 到 12 个月里，我们一直在探索的一件事就是：让 Claude 访问一台计算机，能让它在各方面都变得强大得多；而配置环境，就是让 Claude 访问这台计算机的方式。我已经创建了一个环境，所以就不重新创建了。但如果要从头创建一个新的，我只需点击这个加号按钮。这个 stock watchlist scanner 环境是我今天早上测试时用的环境，你可以看到，只要点击这个写着“启动会话”的按钮就行了。Um, so the first thing that your agent is going to need in order to accomplish anything in the real world is an environment, right? This is a a computer essentially for for Claude to operate on. Um, for the last, oh my gosh, like maybe something like 10 or 12 months at this point, we we've really been about in that giving Claude access to a computer makes it way way way more capable just across the board and configuring your environment is like the the way uh, that you give uh, Claude access to this computer. I've already created an environment. Um, so I'm not going to recreate it, but if I was creating a new one from scratch, I would just hit this four button. But the um, stock watchlist scanner env is the environment that I used when testing things out this morning and you can see that I can just click this button that says start a session.

嗯，所以我对苹果、英伟达、亚马逊、摩根大通、埃克森美孚和好市多感兴趣。我不知道，还有什么应该加进去吗？嗯，嗯。我也不确定。先这样吧，我现在没什么灵感。嗯，那我们按回车吧。这样就会启动一个会话，真正的多智能体编排会在这里进行。嗯，我马上就会讲到这一部分。我们会实际看看正在发生什么，但总的来说，这只是利用我们的 API 目前在托管智能体中支持的现有基础功能，去完成真正的工作。那么，我们来看看会话面板。这里可以看到我在这个平台上做过的其他事情。嗯，抱歉，这里的缩放有点奇怪。嗯，不过我会快速带大家浏览一下，因为我也不确定有多少人在线下实际使用过它。Um, and so I am interested in Apple, Nvidia, Amazon, JPM, XOM, COST. I don't know. Anything else I should add in here? Um, Mhm. I don't know. I'll leave it because I'm not feeling very creative right now. Um, so let's go ahead and hit enter. And what this is doing is kicking off a session. That's where the actual multi-agent orchestration is going to happen. Um, I'll jump into that in just a second. We'll like actually take a look at what's going on, but broadly speaking, this is just using the existing primitives that our API supports today in managed agents and going off and getting real work done. So, let's jump into the sessions pane. This is where you can see my other stuff I've been doing on the platform. Um, and sorry, the zoom is a little funky here. Um, but I'm going to do a whirlwind tour of this because again, I'm not sure how many folks have seen this live in the real world.

嗯，不过在这个例子里，我们有一个代理，也就是我们的股票观察列表扫描代理。这就像是，呃，我想打个比方的话，我忘了那个负责统管整个网络的最后一个人的名字了，不过这就像是金字塔顶端的人，也就是 CEO，负责管理下面所有人的那个。嗯，然后我们有环境，也就是我刚才说的那台 Claude 正在运行的计算机；还有一大堆关于实时动态的信息。嗯，所以你可以看到，这个代理已经启动，正在执行某项任务。嗯，这里展示的就是实际运行过程。嗯，我们马上会更详细地解释其中一些抽象概念，不过基本上，Claude 会依靠自己的自主编排，使用自己定义的一组代理，一步一步地完成这些事情。Um, but in this case, we've got our agent that's our stock watchlist scanner agent. This is like the uh, I guess in the analogy um, I forget the the the name of the final person who's like the overseeing the web of everything, but this is like the person at the top of the pyramid, the the CEO if you will who's managing everybody else underneath them. Um then we've got the environment that's the again the computer that I was talking about where Claude is operating and then we've got a bunch of information about what's happening in real time. Um so you can see the agent is off and doing something. Um so here's here's the actual run happening. Um and we'll explain more in detail about like what some of these abstractions are in just a second, but basically like Claude is going step-by-step based on its own self-orchestration using its own self-defined essentially set of agents to go off and step-by-step achieve each of these things.

所以你可以看到我之前提到的那个会话线程，也就是单个上下文窗口、单个智能体实例，正在并行启动。嗯，确实挺酷的。我也说不好，我觉得挺棒的，基本上这一切都是通过向 Claude 发提示词实现的。我们可以看看之前那个会话里发生了什么，因为它比这个会话进展得更远一些。嗯，这基本上就是在展示我刚才展示的同一件事，不过现在这些都已经完成了。我也不知道，我们的输出文件我就不展开讲太多了，但这里的重点是，我想这次每一步同时运行的智能体大概只有六个。So you can see like the the session thread thing I was talking about earlier which is like the the single context window, the single instantiation of an agent um is is being spun up in parallel. Um so yeah, pretty cool. I I I don't I don't know. I think it's pretty pretty sweet that all this just happens basically by by prompting Claude and uh we we can jump into uh what what uh what's happening in our prior session because it's a little bit farther along than this one. Um so yeah, the this is basically just showing the the same thing I was just showing, but these are all done now and I don't know. Uh I our output file that's I'm not going to go into too much detail on, but basically the the the main point here is that and we've we've had I think in this case it was only like six concurrent agents or so at each step.

嗯，不过 Claude 可以根据任务的复杂程度，动态增减代理的数量。嗯，是的，我想想看还有没有什么想展示的。嗯，我觉得这些基本上涵盖了这个界面的作用。嗯，不过我们当然可以稍后回答问题，进一步深入聊聊。能麻烦你切回去吗？Um but Claude can can spin up and down the the number of agents based on the complexity of the task. Um yeah, I'm trying to think if there's anything else I want to show there. Um I think that that mostly covers just like the the what this surface is. Um but we can of course uh take questions in a little bit to, uh, uh go a little bit deeper. Do you mind jumping back?

主持人：嗯，不，不，我们开始吧。嗯。好。Host: Yeah, no, let's do it. Um. Yeah.

嘉宾：又一次无缝过渡。Guest: Another seamless transition.

主持人：嗯。Host: Yeah.

嘉宾：是的，所以在之前的演示中，Guest: Yeah, so in the in the previous demo,

从控制台到生产环境：调优各个阶段From console to production: tuning each phase

谢谢你，Harrison。嗯，演示得很不错。哇。嗯，这就是一个例子，展示了你如何在控制台中进行原型设计，逐步设置你的流程，并在控制台中测试它。这是很常见的一种做法，可以用来进行试验，看看流程是否按照预期运行。等你对流程的运行方式满意后，无论你的具体使用场景是什么，通常就会把它集成到你自己的应用程序或工作流的某个环节中。比如说，你可能会想按计划触发这个观察名单扫描器，让它每天或每个工作日运行一次；你知道，在开盘前运行，这样我就可以安排它在开盘前五分钟左右运行，从而在开盘前收到报告。thank you, Harrison. Yeah, good walk-through. Woo. Yeah. Um. So, that was all like, uh, an example of how you might prototype in the console and going through and setting up your your your, uh, your flow, testing it out in the console. This is a really common way to go through and experiment and see if the flow is behaving as you expected. And then once you're happy with the way that your flow works, for whatever your use case is, typically you'll go and then integrate this into some part of your your own application or your own workflow. So, whether that's like, oh, I actually want to trigger this uh, watch list scanner on a schedule so that it runs every day or every weekday, you know, before market open so I can like schedule this to run 5 minutes before or something so that I get my report right before market open.

嗯，或者你可能想做这样的事情：从 Slack 或其他类似的界面启动这些流程或触发这些任务。这就是创建 agent 的代码。只要传入这个多 agent 配置，就能启用多 agent。然后，创建第一个事件并把它发送给 lead agent 时，只需说“运行一个工作流”即可。这样就足以让流程启动并正常运行。你只要告诉 Claude：“是的，使用多 agent；是的，运行一个工作流。” 默认情况下，这是最简单的入门方式。不过在这个示例中，Claude 会在执行任务时按需创建所需的子 agent。你也可以在 agent 配置中传入预先配置好的子 agent。Uh, or you might want to do something where you want to kick off these flows or trigger these tasks from Slack or some some other like interface. And so this is, uh, the code for how you would, uh, create the agent. So, it's just like passing in this multi-agent config to enable multi-agent. And then when you create your first event and send that, uh, that first event to the lead agent, you're going to just say, run a workflow. And that's all you need to do to get flows up and running and working. It's just to like tell tell Claude, yes, use multi-agent and yes, run a workflow. Now, by default, this is kind of like the easiest way to get started, but, um, in this example, Claude is going to like on the fly create whatever sub-agents are required to uh to perform that work. And so um it's it's also possible in the uh in the agent config to pass in preconfigured uh subagents.

所以，如果你有一组子代理，知道如何查看 Yahoo Finance、某些新闻来源，或者 SEC 的 Edgar 网站之类的，那么，如果你有专门负责研究流程中重要任务的子代理，也可以在设置代理时将它们配置进去。你可以设置专门负责某项工作的独立代理。还有一点，我觉得现在值得了解的是，你也可以让这些代理使用自定义技能和 MCP 服务器。你可以让它们运行你自己的自定义工具，也就是在环境中运行的工具。所以，它们真的非常强大，而且这一切都能与工作流协同运作。我们还稍微谈到了环境，你也可以在自己的环境中运行它——比如说，使用你自己的沙盒。So, if you have like a set of subagents that know how to like look at uh whatever Yahoo Finance or certain news sources or, you know, the the SEC's Edgar website or whatever. Like if if you have subagents that specialize in certain activities that are important to your research process, you can also configure those uh as part of setting up your agent. So, you can set up uh individual agents that specialize in something. Another thing that's kind of like um really uh I think is is nice to know at this point is that you can also give these agents access to custom skills, MCP servers. You can give them access to uh run your your own um uh yeah, like custom tools that run inside of the environment. So, uh they're really really powerful and then all that works in concert with flows. Um and we talked a little bit about the environment and you can also run that environment on your own um like you can bring your own sandboxes.

所以，如果你愿意，可以在自己的基础设施中运行这个自托管环境。再来回顾一下这个观察名单扫描器的工作方式。通常，如果你要用单个智能体创建一个观察名单扫描器，可能会打开 Cloud Code，然后说：“这些是股票代码。请去研究一下，然后回来告诉我结果。”Cloud 或任何单个智能体都会尽力逐一查看这些股票代码，但只靠一个智能体时，它会在几个方面遇到困难。我们稍后会谈到这些困难，以及 flows 如何解决这些问题。使用 flows 时，我们要做的是把这份股票代码列表拆分开来，然后将每个股票代码分别交给一个扫描智能体。这样我们就进入了扫描阶段。So, you can run this self-hosted environment in your own infra if you want to. So, um just running through again kind of like what that watchlist scanner is doing. So, uh typically, if you were to create a watchlist scanner with a single agent, you might go into Cloud Code and say, "Here's the tickers. Please go research them and come back and tell me." And um Cloud or any any like single agent would try to do its best to look at each of those tickers, but there's a couple different things that agents are going to struggle with when it's just a single agent. And we'll talk about those struggles later and how uh flows addresses those. Um but what uh what with flows, what we're going to do is like split up that list of tickers and for each ticker, we're going to pass it to a scan agent. So, we're going going about like the scan phase.

所以，我们拿一个单独的股票代码交给扫描代理。这是一个隔离的上下文，它只关注这一个股票代码，因此不会受到其他股票代码或其他上下文中搜索结果的影响。扫描完成后，它可以把结果交给下一阶段；此时，审核步骤也处于独立的上下文中，它可能需要，也可能不需要把结果交还给扫描器，或者整理出需要写入报告的详细信息。嗯，我们试过把这种观察列表扫描器作为单个代理运行，也试过将其作为工作流运行。嗯，大多数情况下，当你考虑使用工作流时，可以这样想：“好，我要确保全面、深入地审视要解决的问题，不放过任何细节，而且我愿意为此投入资金。”So, we take one single ticker and pass it to a scan agent, and that is an isolated context where all it cares about is its one ticker, and it can operate without being like influenced by any of the other tickers or any of the other search results from the other contexts. Once that scan is done, it can pass it off to the next phase, and now like the review step is also its own isolated context, which may or may not need to like pass back to the scanner or surface those report details that need to be written out. Um and we went through and we ran kind of like this watch list scanner as a single agent and also as flows. And um in in most cases, when you think about using flows, you want to think about like, "Okay, I want to make sure that I have a really, really comprehensive look at this problem that I'm trying to solve, and I want to go really deep and make sure that I'm leaving no stone unturned, and I'm willing to spend the money to do so."

因此，你在这里右下角看到的是，运行一个流程的成本要高得多；但在中间窗格里，你也会看到，它在找到所有细节方面效果要好得多。如果你用过动态工作流，可能也在 Claude Code 中遇到过类似的情况。嗯，所以，这也是一个机会，可以聊聊你在提示 Claude 构建流程时，可以要求它在不同阶段分配不同资源。比如你可以说：“好吧，我其实想为扫描阶段分配更多测试时的计算资源。”你可以这样提示 Claude：“请使用……嗯，你知道的，或者说，扫描阶段我们可以用 Haiku，这是一款非常快速、高效的模型，它要做的就是上网扫描，对吧？”And so, what you'll see on the bottom right here is that like it costs uh much more to run a flow, but it also in the middle pane, you'll see that it like is much more effective at finding all of the details. And uh if you've used uh you know, dynamic workflows, you've probably experienced something similar with Claude Code. Um so, yeah, it's this is uh also an opportunity to talk about like your um when you're prompting Claude to build a flow, you can ask it to throw different resources at different phases. So, you might want to say, "Okay, I actually want to allocate more test time compute to the scan phase." So, you can prompt Claude and say, "Please use uh um you know, or actually let's say for the scan phase, we can use Haiku, a really fast, efficient model that's all it's going to do is just go out and scan the internet, right?"

所以我们可以说：“你知道，可以针对每只股票启动呃两个 Haiku 智能体，让它们进行一次非常全面的扫描。”这可能是一个成本相当低的阶段。然后你会说：“好，实际上，等我拿到所有网页并开始查看细节时，有时会发现自己漏掉了一些细节。所以，我想在审核阶段投入更多资源。我想让 Fable 深入审核返回的所有细节。所以我会说：‘Claude，请在审核阶段使用一个 Fable 智能体，也许每只股票用两个或三个智能体来审核。’”这样一来，你就可以把这些流程想象成可以根据你希望在流程的各个环节投入多少资源，灵活扩展或收缩的过程。嗯，是的，这是一种非常有效的方式，可以让你在每个阶段调整投入，从而得到你想要的确切结果。And so, we can say, "You know, spin up uh maybe two Haiku agents per ticker to go and like do a really comprehensive scan." And that might be like a very affordable phase. And then you say, "Okay, actually once I get all the web pages back and I'm going through the details, sometimes I'm missing those details. So, actually I want to like invest more in the review phase. So, I want to have Fable review all the details that came back in depth. So, I'm going to say, 'Claude, please use a Fable agent at the review phase and maybe like do two or three agents per ticker at the review phase.'" And so, you can kind of like imagine each of these flows being something that you can kind of like expand and contract based on how much you want to invest in that point in the process. Um and what's Yeah, so this is like a a really effective way to kind of like go through, play around with your investment at each of these phases in order to get the exact result that you want.

所以，嗯，我 definitely 建议你们通过一套全面的评估集来测试，确保这些流程在处理非常重要的大型任务时能按预期运行。我们刚才在控制台里给你们展示的 Watch List Scanner，就是你们可以直接开始使用的模板之一。你们可以直接点进去，一步步操作。嗯，还有一些模板是直接构建并打包在 Claude Code 和 Claude API 技能中的。所以，如果你有 Claude Code，嗯，并且从周日晚上开始就安装了它，那么你应该就能运行 Claude API 托管代理，配合 Watch List Scanner 使用。它可以启动一个模板，让你直接从 Claude Code 管理代理。我觉得这实际上是使用 Claude 托管代理的一种非常有效的方式；总体来说，在 Claude Code 中使用 Claude API 也很有效。And so, um definitely recommend going through like a like a comprehensive eval set to try and make sure that the flows are doing what you want when it's like a very uh you know, big important task. The watch list scanner that we showed you from the console is uh one of these templates that you can start. You can kind of just go, click through. Um there's also some of the templates are built directly and bundled with Claude code and the Claude API skill. So, if you have Claude code um and you installed it starting on Sunday night, uh uh then you should be able to run Claude API managed agents on board with watch list scanner and that can spin up a template um and then allow you to kind of like manage your agents directly from Claude code. I think this is actually like a really effective way to drive Claude managed agents and generally to use the Claude API is from inside of Claude code.

你可以指示 Claude Code，比如说，启动这个资源、启动那个资源，它会使用另一个名为 Ant CLI 的命令行界面来驱动并与 Claude API 交互。所以，很多时候我不会用控制台，而是会用 Claude Code 说：“启动一个能做 X 的托管代理。”那么，今天你们在尝试的过程中，如果不想用控制台，我建议可以从这种方式入手。我们还相当深入地尝试了各种初始系统提示词，或者说，哪种初始消息最适合工作流。我们发现，一开始你越是让 Claude 自己发挥，给它的空间越大，指示越模糊、越自由，它就越能发挥成效。You can direct Claude code to say, you know, spin up this resource, spin up that resource and it will use another command line interface called Ant CLI to drive and and interact with the Claude API. And so, um often times instead of using the console, I'll use Claude code and say, you know, spin up a managed agent that does X. Um so, uh today when you're going through and experimenting, uh this is one of the ways that I would recommend kind of jumping in if you don't want to use the console. Um, and then there's we we experimented quite a bit with like what kind of initial system prompt or, you know, initial message is best suited for flows. And we found that like the the more ambiguous and kind of like free form you let, uh, you are about letting Claude do its own thing in the beginning, the more effective it can be.

所以，你只要告诉 Claude：“嘿，我想让你，嗯，为智能体使用一个单股票代码的工作流，然后，呃，你知道的，去执行这个扫描。这大概就是我想要的。”然后让 Claude 负责搭建初始流程，之后你再稍微微调一下。嗯，那我们来，呃，试试看。你会在屏幕上看到一个二维码。这是用来领取一些 API 额度的，方便你开始使用。如果你还没有 Claude 平台账户，我们本周还宣布了另一件事：如果你订阅了 Max 或团队方案，订阅中会包含一些 API 额度。所以，嗯，你也可以关联自己的订阅，确保能用上这些 API 额度。如果你有 Max，我想，呃，Max 5X 每月能给你 100 美元的额度。So, just tell Claude, "Hey, I want you to, um, use a workflow one ticker for agent and, uh, you know, go out and do this scan. This is what I'm kind of looking for." And let Claude take control over like building out the initial flows and then you can kind of like, uh, fine-tune that afterwards. Um, so let's, uh, let's play around. So, on screen you'll see a QR code. This is for, um, uh, some API credits so you can get up and running. If you don't already have a Claude platform account, another thing we just announced this week is that, uh, with Max and team subscriptions, you, uh, you get some like API credits that are part of your subscription. So, um, you can also like go link your your subs to make sure that you have those API credits. So, if you have Max, I think it's like, uh, Max 5X, you'll get like $100 a month.

嗯，然后 Teams 的上限是 500。接下来大概再花 50 美元，就可以去尝试一下 Claude 托管代理流程。嗯，好，我觉得我们现在可以先暂停一下。我们还有一些内容想继续讲，不过我们先暂停吧。请大家试试看，然后我们可以开放提问，如果有人有问题的话。Um, and then Teams is up to 500. And then this is like another 50 bucks to go and, uh, experiment with, uh, this Claude managed agents flows. And, uh, yeah, I think at this point we can pause. We have like a bit more content that we want to go through, but what let's pause. Please try it out and then, um, yeah, we can we can open up the floor for questions, uh, if anybody has

问答环节，第1轮Q&A, round 1

一开始有什么问题吗？好，嗯，这是个好问题。问题是：当你在 Claude Code 中运行 Claude Code 里的 {slash} Claude API skill 时，它是在我的机器上运行，还是作为 Claude 托管代理的一部分运行？对吧？Claude API skill 和 Claude Code 会根据你的 Claude Code 配置，在你的机器上或云端运行。这和 Claude 托管代理完全是两个独立的系统。所以，你的 Claude Code 会根据配置在本地或云端运行。然后，当你启动它时，它会从 Claude Code 环境发起 API 调用，以创建代理、创建会话，并在 API 中创建所有这些不同的资源。any questions off the bat. Okay, so, uh, good question. So, the question was, when you're in Claude Code and you run the {slash} Claude API skill in Claude Code, is that running on my machine or is that running as part of Claude managed agents, right? So, um, the Claude API skill and Claude Code are going to run it on your machine or in the cloud depending on how you have Claude Code set up. And that is like a completely separate system from Claude managed agents. So, you have you'll have like your Claude Code is going to be running either locally or in the cloud depending on how you have Claude Code set up. And then, when you kick that off, that's going to make API calls from the Claude Code environment to create the agent, create the session, create all of these different resources in the API.

呃，有关于它如何与插件合并的路线图吗？Uh is there a roadmap for how it merges with plugins?

主持人：所以，目前插件涵盖了很多不同类型的资源，对吧？其中一种就是技能，对吧？嗯，所以你看 API 的接口时，会注意到里面有一个叫 resources 的字段，另外还有一些与技能相关的 API 字段会出现在插件中。所以，简而言之，插件的各个组件现在就应该可以直接接入 API 接口。至于把它合并进去……不好意思？哦，没关系。另一种情况是，你的工作流生成了某个标量之类的东西，然后你可以把它拉回 Claude Code 里？或者你介意再多说一点吗？明白了。好，所以问题基本上是，假设你创建了一个动态工作流，在这个例子中是用于研究潜在的收购对象，那么你能不能把这个工作流作为插件有效地分发出去？Host: So, plugins right now encompass a bunch of different like types of resources, right? One of those is is skills, right? Um and so, when you take a look at the API surface, you'll notice that there's there's a field called resources associated with it, and then there's also like some other API fields that are related to things like skills that show up in plugins. And so, the the short answer is that like the individual components of a plugin should just drop into the API surface already today. As far as like merging it into Sorry? Oh, no worries. So, the other direction being like your workflow generates a scalar something, and you can pull it back into your Claude Code? Or Do you mind just saying a little bit more about that? I see. Okay, so the question is basically given a dynamic workflow that you've created for, in this case, um research on potential acquisitions, can you distribute effectively the workflow uh uh as a plugin?

嗯，答案是现在还不行。至于未来某个时候是否会开箱即用地得到全面支持，我也没有确切答案。不过，你可以把工作流运行过程中生成的提示词拿出来，直接放进 Claude Code 之类的工具里，然后提交进去。它是个文本文件，效果可能会非常接近。对，对。所以，问题是，能不能在 Mac 环境或其他环境中运行一些代理，而不只是运行 Claude 托管代理的那个托管环境？嗯，我们有一种叫作自定义工具的抽象机制，基本上就是让模型向你发起调用，请你执行某个操作。具体来说，你可以在工作流运行过程中，为你的代理定义一个自定义工具。Um The answer is not today, and I don't have a good answer for if that will be like fully supported out of the box at some point in the future, but you could take the prompt basically that's generated as part of your workflow runs, and just drop that into like Claude Code or wherever, um check it in, it's a text file, it'll uh probably be pretty pretty close. Yeah, yeah. So, the question is, what about running like some agents in like a Mac environment or other environments that aren't just like presumably the hosted surface where uh Claude managed agents are running? Um we have this abstraction called custom tools that's basically like a way for the model to kick out to you to ask you to take some action. So, the way that this would work is you would define a custom tool uh for your agent as part of the workflow run.

然后，任何时候你想在比如一台本地 Mac OS 机器上，或者互联网上其他任意一台计算机上运行某个东西，你基本上只需要调用那个自定义工具，在那台机器上执行你想做的操作，然后把结果返回给模型。这样回答了你的问题吗？And then anytime you want to run something like on, let's say, a local Mac OS machine or like some, you know, arbitrary other computer on the internet or something, you would basically just like take that custom tool call, do whatever you want to do on that machine, and then return the result back to the model. Did that answer your question?

嘉宾：好，太棒了。Guest: Okay, sweet.

主持人：是的，完全同意。所以问题是，我们迟早都会在意 token 消耗，需要随着时间推移优化工作流，避免耗费大量 token。为此有很多不同的方法。这其实是个可泛化的问题：根据你内部关注的标准，在某个提示词上进行爬山式优化，无论标准是 token 消耗、能力表现等等。所以，嗯，我不记得我们有没有发布过这个爬山优化技能。Host: Yeah, totally. So, the question is we all care about token spend at some point and we need to be able to optimize our flows over time to not basically spend tons of tokens. So, there are a bunch of different ways to do this. This is kind of a generalizable problem that's like hill climbing on some prompt based on whatever criteria it is that you internally care about, whether it's token, whether it's capability, etc. And so, um there is I forget if we published the hill climb skill

嘉宾：是的。Guest: Yeah.

主持人：在 Cloud Code 中，Cloud Code 里有一个爬山优化技能，应该至少能帮你优化提示词本身。至于像流程操作这样的调优，嗯，这么说有点像是在敷衍，但实际上，你会拥有 Cloud 在流程运行过程中执行所有操作的完整记录。你可以直接调出这些记录，交给 Cloud，让它找出流程中可能做得过多或不够的地方，它基本上就能自我优化。嗯，这样你就可以让 Cloud 根据你想要的任何标准进行优化。不过我想说的是，我们是在周五推出这个功能的，我猜这种原生集成会是很多人非常关注的重点。Host: in Cloud Code. So, so there's a hill climb skill within Cloud Code that can probably help you hill climb at least the prompt itself. As far as the tuning of like the the flow operations, um this is kind of a cop-out answer, but you effect you'll have the full transcripts of everything that Cloud has done as part of the the flow runs. And so, you can just pull those transcripts, give them to Cloud, ask Cloud to identify any like places where the flow is maybe doing too much or not doing enough, and it's able to essentially optimize itself. Um and that way you can ask Cloud to optimize based on whatever criteria you you you want. But, I will say like we shipped this on Friday and I suspect that this type of like native integration is something that is going to be top of mind for a lot of folks.

嘉宾：是的，我还想补充一点，还有一个叫 Dreams and Dreaming 的产品。在 Cloud 托管代理中，你可以指定一组会话，让它们经过这个“做梦”流程，从而对与这些会话相关的记忆库进行提炼和改进。你可以挂载一个记忆库，运行工作流，然后在运行了 10 次左右之后，让这些会话经过 Dreams 流程，让它们“做梦”，以便提炼记忆，并逐步改进不同会话之间的记忆。所以我不知道你对 Dreams 还有没有什么想法，或者……Guest: Yeah, one thing I would add to that, too, is there's another product called Dreams and Dreaming where you can within Cloud managed agents give a set of sessions to go through this dreaming process where it can kind of like crystallize and like basically like improve the memory store that is related to those sessions. So you can attach a memory store, run your flows, and then after you've done it, you know, 10 times or whatever, you can have those sessions run through dreams and have those dreams kind of like crystallize and improve the memory from from like session to session. So I don't know if you have uh more dreams thoughts or

主持人：想法，但是Host: ideas, but

嘉宾：是啊，我想我基本上把手动打造自己的 dreaming 描述了一遍。不过我忘了，我们有自己的产品可以帮你完成这件事。Guest: Yeah, I think I basically described hand rolling your own dreaming. But I forgot we have our own product that does this for you.

主持人：是的。对。所以你可以、你可以使用产品构想，也可以使用流程来实现构想。所以……Host: Yeah. Right. So you can you you can either use the product dreaming or you can use flows to implement dreaming. So

嘉宾：对，完全可以。所以如果你只是想……抱歉，我先退一步。问题是：在流程运行时，你能不能直接在控制台里查看多智能体的拓扑结构？可以，待会儿我们会给大家机会亲自看看。嗯，不过如果你还想围绕它构建自定义 UI，也完全没问题，因为这些流程中的每一个都会启动子智能体，而它们都有各自的会话线程。这些线程就是它们运行所处的上下文，而每个线程的所有信息都可以通过 API 获取。所以，你可以根据自己的需求，围绕它构建各种有意思的 UI。嗯，我看 Jay 可能正准备……Guest: Yeah, totally. So you can you can visualize it directly in the console if you just want like sorry, stepping back. The question is can you can you visualize the the multi-agent topology as the flow is running? Within the console you you can and we'll we'll give you guys a chance to check it out in just a second. Um but if you're if you want to build like a custom UI around this thing as well, you totally can because each of these flows spins up sub-agents that have their own session threads. That's like the context in which it's operating and all of that information on a per thread basis is accessible via API surfaces. So you can spin up arbitrarily interesting UIs around it for for whatever it is that you want. Yeah, I see Jay's probably about to

主持人：是的，你可以喜欢、】【Host: Yes, you can like

嘉宾：是的。Guest: Yeah.

主持人：看看所有线程。你可以深入查看各个子代理，也可以看到它们在做什么。如果你通过编程方式访问这个程序，还能看到一大批非常全面的事件，这些事件会在会话流式传输时发出。因此，你会看到每一次工具调用及其结果，也能看到何时从子代理交回给主代理，也就是负责统筹的代理，等等。所以，目前在这个流程中，仪表板的 UI 显然还有改进和迭代的空间。不过，你完全可以深入查看这里的各种详细信息。比如，你可以看到这是一条“收到线程消息”事件。这样你就能知道，主代理实际传给子代理的内容是什么。Host: Look at all the threads. You can drill into individual sub-agents. You can see what they were doing. You can also look at if you're you're accessing this programmatically, there is a huge like really comprehensive set of events that are emitted like as the session is streaming and so you'll see every single like tool call and result and like when it's handing off between from from the sub-agent back to the main uh, know, the lead agent, etc. So, like, um, yeah, within the flow right now, the UI in the dashboard is definitely something that we'll continue to improve and iterate on. Um, but, yeah, you can you can absolutely dig in and see like lots and lots of detail here. Um, so, you can see like, okay, yeah, this is the thread message received event. And so, you can tell like, okay, what did the lead agent actually pass to the sub agent?

你可以在这里查看该负载，从而深入了解整个图中每次转换具体发生了什么。And you can look at that payload here to deeply understand kind of like, what is actually happening for like every transition in this entire graph.

嘉宾：我还想说，这也是基于公共 API 构建的，就像仪表板本身一样。所以，你在这里看到的任何东西，都完全可以自己动手搭建。问题是，它是否能意识到自己的预算？目前，模型在运行时并不知道自己的预算。预算主要是作为护栏，确保它不会偏离轨道，呃，导致 token 花费失控。嗯，我们发现，至少在目前最新的模型中，如果你告诉它预算，嗯，Claude 可能会有点焦虑，而我们不，呃，我们不希望这种情况发生。嗯，不过我很相信，随着时间推移，这方面会有所改善。Guest: And I'll also say this is built on the public API, like the dashboard itself. So, anything that you see here, you could totally hand roll yourself. The question is, is it self-aware of its own budget? So, right now, the model does not know its budget like as it's as it's running. The budget is primarily like guardrails to make sure that it doesn't go off and uh, run away token spend. Um, we found that at least in today's latest models, um, if you tell it the budget, um, Claude can get a little bit anxious and we don't uh, we we don't we don't want that to happen. Um, but I would highly suspect that over time that that will improve.

主持人：所以，flows 的工作方式是，如果你用到 5 美元，就会收到一个事件，大致意思是“已达到预算上限”。然后，如果你想看看发生了什么，并且可能提高预算，只要会话还在运行，就可以这么做。如果你想再增加一些资金，也可以，flow 会从上次停下的地方接着运行。所以这里有一个很棒的暂停机制：如果你提高了预算，已经完成的工作不会被丢弃。嗯，不过这完全由你自己决定。对。所以，问题是，计划如何与 Claude Tag 集成？嗯，Claude Tag 本质上是一个 Slack 机器人，可以把你的所有信息整合到 Slack 界面中。所以，它属于一个稍微不同的产品界面；而这是一个 API 界面，如果你想直接基于 Claude 构建集成，就可以使用类似这样的东西。Host: So, the way that it works with flows is if you hit $5, you get an event basically being like, budget reached. And then if the if you want to like take a look at what happened and potentially up the budget, you can as the session is running. And if you want to add more money, then you can and the flow will pick off like where it left off. So, there's this really nice pausing component where any work that you've already accomplished is not thrown away if you up the budget. Um, but that's totally up to you based on your own Yeah. So, the question is, what is the plan for integrating with Claude Tag? Um, so, Claude Tag is essentially like a Slack bot that connects all of your information, uh, uh, together within the Slack surface. And so, it's a slightly orthogonal product surface, whereas this is an API surface where if you want to like build an integration with Claude directly, you would use something like this.

嗯，至少目前我觉得应该不会有任何直接连接，其他……你知道吗？对，对。如果你想做个类似“给自己打标签”的东西，可以用流程。是的，如果你在用……[嗤笑]没错，没错。所以，嗯，对。对。Um, I doubt at least for now that there would be any direct connection other You know what? Yeah, yeah. If you wanted to like build something like tag yourself, you could use flows. Yes, if you're using [snorts] Exactly, exactly. So like, yeah. Yeah.

嘉宾：是的。我还想说一点，就是，嗯，我们有很多方面都想做得更好，比如让所有这些产品界面更好地协同工作。所以我想说，接下来几个月左右，大家可以期待看到这些衔接处逐渐消失。你可能已经注意到，我们把 co-work 和 chat 中的一些功能合并了。现在你可以在很多不同的界面中使用 Cloud Code。随着时间推移，我们会逐步缩小这方面的差距。还有，嗯，Cloud Tag 确实是使用很多这些工具的一种非常有效的方式。我们使用它的频率高得有点出乎意料。对我们来说，Cloud Tag 是很多实现工作中每天都会用到的重要工具。所以……Guest: Yeah. I I would say the other thing too is that like um this is this is something like we want to improve on a lot is like making all of these product surfaces work better together. And so I would say expect, you know, over the next you know, several months or whatever to continue to see these seams start to like disappear. You may have noticed that like, you know, co-work and chat kind of like we merged some of those. You can see now you can use Cloud Code from lots and lots of different surfaces. So over time we'll kind of like uh close the gap there. And yeah, um Cloud Tag is definitely like a really effective way to to use a lot of these tools. And in like we use that a surprising amount. Cloud Tag is like a huge daily driver for us for for a lot of implementations. So

主持人：所以问题是，智能体之间有没有办法互相通信，比如流程中的子智能体？嗯，你不能通过 API 直接告诉子智能体 A 给子智能体 B 发消息，并发送载荷 X。不过，所有子智能体都共享一个文件系统。因此，如果工作流中需要它们互相通信，它们实际上可以直接利用文件系统来回发送消息。但除了这些“智能体消息已发送”和“智能体消息已接收”的记录（你可以在其中看到消息来回发送）之外，并没有直接的抽象机制，让你强制一个子智能体给另一个子智能体发消息。是的，所以问题是，流程能不能以某种方式修改自身，并根据可能发现的问题逐步优化自己的提示词？Host: So the the question is is there a way for the agents to communicate with each other um like the sub-agents within the flow? Um there's not like a direct like uh uh way for you to like via the API say sub-agent A like message sub-agent B and like send payload X. Um but the the sub-agents all share a file system. And so if there is a need uh within your workflow for them to communicate with each other, they can effectively just use the file system as a way to like send messages back and forth. Um but there's not like a direct abstraction that uh other than these agent message sent and received where you can like see the message sends back and forth. There's not a direct way for you to like force one of the sub-agents to message another one. Yes, so the question is is there a way for the flows to essentially modify themselves and hill climb their own prompts based on issues that they might find.

嗯，最好的做法可能和我们刚才讨论的 hill climb scale 类似，也可能会用到 dreaming 产品界面，这样智能体就能看到自己过去做过什么，并从错误中吸取教训。目前还没有这样的东西，呃，事实上我直接快速演示一下。我们有一个 deployment 的概念，基本上 deployment 就是按计划运行这些会话的一种方式。你没法按计划让它自己不断改进，不过我完全可以想象这样一种情况：你可以用这些组件来构建这样的功能，那会很酷。Um so the best way to do this is probably similar to like the hill climb scale that we were talking about just a second ago, as well as potentially the the dreaming product surface, so that the the agent can see like what it did in the past and learn from its mistakes. There's not like such a thing as um so actually here I'll I'll just show really quickly. We we have this um concept of deployment, and basically deployments are um a way to run these sessions on a schedule. And you there's not a way for you to like on a schedule like self-improve the thing, but I could totally imagine a world where you like you could build that using these components, and that'd be pretty cool.

嘉宾：还有，比如我们一直在提的 hill climb 技能，它也是 Claude API 技能的一个子命令。嗯，如果你在 Claude Code 里，可以输入 / Claude API hill climb，然后告诉它你想通过迭代改进什么东西。这有点像是把 Andrej Karpathy 的自动研究理念融入一个人人都能使用的技能里，让大家可以迭代改进自己的提示词。比如，你可以先从一个现成的提示词开始，也就是那个用来扫描观察名单的提示词，然后把它拿到 Claude Code 里，说：“Claude API hill climb 这个提示词。”接着它就会和你一起迭代，弄清楚我们究竟要优化什么，然后不断打磨，直到为你生成一个很棒的提示词，让你可以拿去满足自己的需求。Guest: Also like for the the hill climb scale that we keep mentioning, this is another sub-command of the Claude API skill. Um so if you're in Claude code, you can do like {slash} Claude API hill climb, and then tell it what thing you want to like iteratively improve on. This is kind of like a uh taking um Andrej Karpathy's auto research and sort of baking it into a skill that everybody can use to kind of like iteratively improve on their own prompts. And so if you start with a prompt that you know, the one off the shelf to do the watch list scan, then you can uh take that prompting and go into Claude code and say, you know, Claude API hill climb this prompt, and then it will work with you to iteratively like figure out, okay, what are we optimizing for, and then it'll just kind of like grind until it gets like a really nice prompt for you to then like use that uh to to yeah, meet your needs.

嗯，所以这取决于爬山过程中发生了什么，有时它会给你……这真的完全取决于你要解决的具体项目或问题。嗯，几周前 Haiku 发布时——或者说，一切发生得太快了，我想可能是上周吧，我也不确定。嗯，Haiku 发布后，我做了一个可以和孩子们玩比手画脚游戏的工具。它就只有一个摄像头，会给我们一张卡片，我们试着把卡片上的内容表演出来，然后 Haiku 会猜我们在做什么。于是我用了这个爬山命令，Claude 为我生成了一个产物，让我可以录下自己表演这些搞笑动作的过程，比如划船、打网球之类的，这样它就能改进并迭代提示词，还会指出：“哦，我们需要这么多帧每秒，还需要更新提示词来实现这个。”Some So depending on uh what happens during the hill climb, sometimes it will give you it really depends on exactly the project or like the problem you're trying to solve. So I uh I when Haiku was announced a couple weeks ago, or everything happened so fast, I think it might have been last week, I don't know. Uh when Haiku was announced, I built this tool for playing charades with my kids. So it just has like a webcam, and it'll give us a card, we try to act it out, and then Haiku will guess like you know what we're doing. And so I use this the hill climb um command and Claude generated an artifact for me that let me record sessions of myself doing like these hilarious things rowing a boat and like playing tennis and all this so that it could like improve and iterate on the prompt and say oh we need this many frames per second and we need to like update the prompt to do this.

因此，具体要看问题是什么：运行 hill climb 时，Claude 可能会也可能不会提供一个人工界面，让你可以针对它反复迭代；你也可以让 Claude 替你完成这件事。所以它其实相当强大。And so depending on what the problem is like when you run hill climb Claude may or may not give you an a human interface to kind of like iterate on that and you can also like request Claude to do that for you. So it's actually like quite powerful.

主持人：是的，这个问题涉及纳维－斯托克斯方程，以及这些突然间接连涌现的科学发现；像这样的东西，是不是我们会用来实现这一目标的工具。我觉得，也许在我们的演示中，股票行情代码这个例子还不够有野心，因为你确实说到点子上了：推动科学取得重大进展，基本上就是缩短完成这类工作的时间。嗯，流（flows）基本上是一种在测试时计算中投入算力来攻克难题、让模型更快获得更强能力的方法。至于我们内部运行这类研究工作负载时具体采用什么样的框架，我不便谈细节；但我们认为，尤其是通过 API 接口开放这项能力，是让公众普遍能够做出这类发现、取得科学进步的第一步。Host: Yeah so the question is about Navier-Stokes and like these scientific discoveries that are all of a sudden happening kind of all at once and if something like this is like what we would use to to to to do to do that. I think maybe in our our demos our our stock stock ticker is not ambitious enough because you pretty you pretty much hit the nail on the head right that like the way that you pull forward substantial amounts of scientific progress is basically shorting the shortening the timeline that it takes to do this kind of work. Um and flows are basically a way excuse me to throw a test time compute at hard problems and get more capability faster out of the models. And so um I can't speak to like specifics of like how internally we like we what our harnessing looks looks like for running these types of research workloads but like we think especially as an API surface that unlocking the power of this is like the first step for us making these types of discoveries or um scientific progress like broadly doable by the general public.

所以，嗯，是的，我对此相当兴奋。So um yeah I'm pretty jazzed about that.

嘉宾：对，我还想补充一点：在我们之前为 Claude 托管智能体搭建的多智能体系统中，如果你有一个协调器和一组智能体，那么能够启动的子智能体数量会受到一些限制。而现在使用 flows，我想可以启动 1,000 个智能体。所以，另一个有意思的地方是，你的主智能体可以同时启动多个 flow，也可以等一个 flow 完成后再启动另一个 flow。这样一来，你就可以把一个规模非常庞大的项目交给一组智能体来处理。其实我本来真希望它现在已经完成了，但还没有。周五我启动了一个由 1,000 个智能体组成的 flow，让它们构建一个操作系统给你们看。它已经运行了大约 85 个小时。Guest: Yeah one other thing that I'll add to that too is that like in our previous multi-agent setup for Claude managed agents if you where you have like a coordinator and then a roster of agents we had some limits around like how many sub agents you can spin up. With the flow with flows right now, I think it's 1,000 agents. And so like the other the other like interesting piece of this is that your lead agent could spin up multiple flows at a time or like after one flow finishes, start another flow. And so you might have like a really massive project that you want to like give to give to a set of agents. I actually like was really hoping it would be done by now. It is not. But I I on Friday kicked off a flow with 1,000 agents to build an operating system to show to you. And it has already been working for like 85 hours.

所以它还没做完，现在大概进行了三四个流程，而且一直在这个更大的项目上和一大堆智能体协作。所以这些事情你完全可以交给它做。显然，因为它们要花的时间稍微长一点，这个项目我好像是周五才刚开始，然后就没管了，想着也许到周一早上 9 点它就能准备好。不过这绝对是那种你可以交给它做的项目，比如云管理智能体流程。对，这个我们留到之后再说，不过它实际上非常轻量。我当时说：“嘿，Ken Thompson 一周就做出了 Unix。你能不能试着做一个兼容 Unix、带 Python 解释器的操作系统？用流程来做，大胆一点，最多可以用 1,000 个智能体。”And so like it's not done yet and it's like three or four flows in and has been working with like tons and tons of agents on this like bigger project. And so these are things you can absolutely throw at it. And obviously because they take a little bit longer, this is something that you I kind of like just started on Friday and then walked away and was hoping that maybe by 9:00 a.m. on Monday it would be ready, but definitely like the type of project that you could um you could throw at yeah, cloud management agent flows. Yeah. Let's save that for an after, but yeah, it was like it was actually like very lightweight. I was like, "Hey, Ken Thompson made Unix in a week. Can you try and make an operating system that's Unix compatible with a Python interpreter? Use flows, you know, feel free to be brave and use up to 1,000 agents."

基本上就是这样。然后就让它运行，我还给了它一个 GitHub 仓库，让它可以在那个仓库里干活。我上次查看时，那个 GitHub 仓库里大概已经有 1,000 次提交了。所以，是啊，这些智能体还在埋头干活。嗯。And that was basically it. And then like let it run and I also gave it a GitHub repo so that it could work in that GitHub repo. The last time I checked there was like 1,000 commits in the GitHub repo. So it's yeah, the agents are still grinding on it. Yeah.

主持人：呃，这个问题是，我们有没有注意到，像是以同样的方式不断增加子代理，会出现收益递减？就像一个组织随着规模扩大，每增加一名新员工，协调成本之类的开销都会随之增加。这确实是个很好的问题，但我没法给你一个准确的数字或答案。嗯，我觉得整个研究领域都还很新，所以如果未来能够更精确地量化这类问题，我一点也不会意外，比如说，衡量每个子代理的能力曲线，或者在增加一个额外代理时，边际增益的斜率。嗯，不过我现在还没有一个靠谱的数字。嗯，对，这是个好问题。Host: Uh the question is if we've noticed diminishing returns from like spinning up more sub-agents in the same way that like an organization like on an incremental basis for every new employee you add at a certain size, there's like coordination costs and such that that increase over time. That's actually a really good question that I don't have a perfect number or answer for you on. Um I think that like this whole research area is pretty new and I would be very unsurprised if like this these types of exact like uh numerical assignments to like or basically like the it's almost like the capability curve per subagent or like the incremental you know, the slope of the line as you're adding a new an extra agent um is more calculable in the future. Um but I I don't have a good number right now. Um yeah, that's a good question.

嘉宾：是的，我想说的一点是，使用流程的好处在于，你可以有条理地让它们处理问题，而不是毫无章法地一股脑儿丢过去。比如说，如果让同样的 100 个智能体来解决问题，使用流程会做得更好，因为通过门控和不同阶段，这些上下文之间有一定的组织性和隔离性。嗯。Guest: Yeah, one thing I will say is like with flows, you get the benefit of like throwing them at a problem in an organized way versus an unorganized way. So like if you take the same 100 agents and throw them at the problem like it's going to do better with flows because there's some organization and isolation between those contexts with gating and these different phases. So.

主持人：那么，问题是有没有办法防止无限循环或错误级联？答案是有，某种程度上可以。Host: So the question is is there a way to prevent infinite loops or cascading errors? And the answer is yes, sort of.

嘉宾：嗯，确实有一些机制可以确保会话不会失控。因此，我会设置一个预算，对吧？具体取决于你的预算是多少。不过除此之外，Claude 在检测行为异常的子代理并相应地管理它们方面其实做得相当不错。嗯，我们发现，至少在我们自己的评估中，Claude 很擅长检测这类错误，并自动将它们剔除。我们刚才还谈到了“做梦”、爬山搜索之类的东西，以及从过去的运行中学习。我们认为，如果你想非常有针对性地优化工作流中的某些部分，或者优化 Claude 正在尝试完成的某些任务，这确实是一种非常有效的方式。Guest: Um the there's mechanisms to make sure that like you don't end up with runaway sessions. So for that I would throw on a budget, right? Based on what whatever it is that you uh your budget might be. But then otherwise Claude is actually pretty good at detecting subagents that are misbehaving essentially and managing them accordingly. Um and so we found that at least in our own evaluations like Claude is Claude is pretty good at like detecting these types of errors and and cutting them out um uh automatically. The other thing is we were talking about like dreaming and hill climbing and stuff and like learning from past runs and we think that that's a really powerful way to if you like very specifically want to optimize parts of your flow or parts of what Claude is trying to accomplish, um that's like a really really nice way to do so.

让我复述一下你的问题，看看我理解得对不对。嗯，基本上你是想弄清楚，是否有办法根据某些标准，比如说较小的一组内容——也许像镜头切换这样的信号——只执行某个流程中的部分步骤。基本上，你不想因为没发生什么事情而白白消耗 token。对吧？目前平台上还没有真正类似的功能，除了让 Claude 在单个流程中有条件地执行某些操作。坦率地说，如果是我，我可能会在你的应用层里，在 flows 上面再叠加一层东西，用来进行这种检测。嗯，这样你就可以大致确定要执行的工作量了。Let me state your question back to you and tell me if it's if it's correct. Um so basically you're you're trying to understand if there's a way to like based on some criteria like a smaller set may maybe be a signal like a shot change, you should only execute potentially like some parts of a flow or some some parts of a like like basically you don't want to burn tokens because like nothing happened. Right. There's not really anything like that on platform today other than just like telling Claude to conditionally do stuff with a single flow. Like if I had to be totally honest with you, I would probably like layer something else on top of flows yourself within your application layer to make that detection. Um so you could like size basically the uh the the the work that you're trying to do.

我觉得，对于那种单个智能体反正 30 秒就能完成的事情，工作流可能不太合适。嗯，它更适合那些雄心勃勃、复杂的任务。所以，是的。没错。当然。问题是，呃，我能不能稍微聊聊智能体之间的上下文护栏，以及每个智能体各自拥有的世界观。嗯，每个智能体都会在自己的会话线程中启动；从数据库层面看，它们在物理上是完全隔离的，拥有各自独立的上下文窗口。因此，默认情况下，这些上下文窗口不会通过文件系统之类的途径相互泄漏。除此之外，每个智能体都有自己的系统提示词。所以，无论你是想使用自己编写的系统提示词，还是让 Claude 生成系统提示词，嗯，都可以；这是确保各个角色彼此完全独立的一种非常有效的方式。I don't think that flows would be a good fit for something that's like, you know, single agent can accomplish in 30 seconds anyway. Um it's more for like these ambitious, complex tasks. So, yeah. Yeah, for sure. So, the question is uh if I can talk a little bit about the context guardrails between the agents and effectively like the world view that each of the agents have. Um so each of the agents is spawned into its own session thread that's like totally like physically in the database like separate context windows. And so there's there's no leakage of those context windows via like file systems or anything like that um by default. Um beyond that, each of the agents has their own system prompts. So, whether you want to bring your own system prompts or just have Claude spin up the system prompts, um you can do so and that is like a really good way to enforce the totally separate personas.

另外，线程消息的发送和接收是了解智能体之间通信情况的好方法。如果你想确保中层管理智能体没有对它的下属做一些你不希望它做的事，就可以查看这些消息。总的来说，这就是它的运作方式。回答了你的问题吗？ 我明白了。这个问题是关于我提到的共享文件系统——它可以作为子智能体彼此交互的一种方式，以及有没有办法在物理上将它们隔离开来。如今要做到这一点，基本上就是通过提示来实现。如果你希望智能体只关注某一组非常具体的目录路径，或类似的工作范围，只要在提示中说明即可。Um on top of that, um the I mean the thread message sends and thread message receives are a good way to see the inter inter-agent communication if you want to make sure that like, you know, middle management agent isn't like doing something with uh its underlings that that you don't want it to do. Um but generally speaking, those are like that's how it works. Did that answer your question? I see. So, the question is about the shared file system that I alluded to as like a methodology for sub-agents interacting with each other and if there's a way for for to like physically separate those out. Um the way to do that today would basically just be prompting um, and telling like if you're if you want your agents to only focus on like a very specific set uh, like a subset of directory paths or something for where it does its work, um, you would just prompt that.

目录路径没有类似权限控制层的东西，不过说实话，这看起来相当实用。嗯，也许我们会把这个想法带回去，再回来给出更好的方案。所以，后续的问题是，能不能用类似 hooks 的东西来控制对文件系统不同部分的访问？在托管代理中，这个功能的实现方式不完全像 hooks，不过你其实在某种程度上已经替我回答了这个问题，因为实现它的一种方法是使用一种叫作 tool permissions 的抽象：每当代理想要读取文件时，根据你提供的配置，它可能需要向你，也就是向你的系统，请求许可。在这种情况下，如果你想锁定 /mount/fubar，那么你只要读取这个工具发出的请求——也就是 Claude 发出的读取该文件的请求——检查实际运行该读取请求的代理，如果你不希望它获得访问权限，就直接拒绝即可。There's no like permissioning layer for the directory paths, um, but honestly like that seems pretty useful. Um, maybe we'll uh, maybe we'll take that back and come back come back with something better. So, so the the the follow-up was what about using something like hooks to uh, control access to different parts of the file system? Um, that's the the way that this would work in managed agents is not quite like hooks, but you're actually kind of answering the question a bit for me because um, one way that you could do this is we have an abstraction called tool permissions that basically like anytime an agent wants to read a file, it depending on the configuration that you provide, might need to ask you for permission or you being like your system for for permission. And in this case, if you wanted to like lock down, you know, {slash}mount{slash}fubar, um, then you could just read the tool basically the request that Claude is making to read that file, check the agent that's actually like running that uh, read request and just deny it if if you don't want the agent to have access.

哇，谢谢，真的很感谢。哦，天哪，你这是在给我铺垫话题啊。嗯，我们接下来要……所以，这个问题是关于凭证管理的。嗯，这里的重点是：“天哪，一千个智能体，这也太疯狂了。这么放任它们在互联网上运行，怎么看都不安全。”嗯，我们基本上会在这场演讲的后半部分，整整花一半的时间来讨论智能体安全。问题是，能不能把人工反馈纳入工作流程。如今实现这一点的方式是使用所谓的自定义工具。基本上，就是给 Claude 一个用于请求人工反馈的工具；需要时，Claude 就会向你请求反馈，你可以通过调用这个工具提供任何你想给出的回复。Wow, thank you. Appreciate that. Oh man, you're you're teeing me up. Um, we're going to So, the question is about uh, credential management. Um, so like the the the punchline here, right, is like, "Oh my gosh, like a thousand agents, that's crazy. Like that doesn't seem like safe to just like let let run wild on the internet." Um, and we're going to spend like basically the entire second half of this presentation talking about agentic safety. The question is if there's a way to allow human feedback as part of a flow. The way that you would implement this today is using something called custom tools. Basically like give Claude a tool that's like request human feedback and then Claude would request that human feedback from you whenever it needs it and you could provide whatever response you wanted and from that tool call.

所以，问题是，嗯，以构建操作系统为例，如果它挂掉了怎么办？然后，呃，怎么查看中间状态？在它运行的过程中，你怎么和它交互？你想回答这个吗，因为……So, the question is um, in the example of building an operating system, uh, what if like what if the thing dies or something and like like how do you see the intermediate states? How do you interact with it as it's going? Do you want to take that since

主持人：是的，当然。我在提示词里有一部分是让它组建一个文档写作团队，我还说用 Sonnet 5.5。我很喜欢它生成的文档。每个流程结束时，都会有一个文档写作团队回过头来，撰写关于该阶段实现内容的文档。这样我就可以直接去 GitHub 仓库，边做边查看文档。这是一种做法。另一种做法——你刚才问过能不能暂停或停止它，对吧？还有一件事我们稍后会谈到，就是中断会话：你可以中断流程，然后说：“其实你走错方向了，请马上停下来。”你也可以提示 Claude 生成报告，或者在进行过程中更新某份报告。通常它会给你一些 Markdown 文件，你可以打开查看。Host: Yeah, sure. I so um part of the prompt I asked it to have like a a team of docs writers and I was like use Sonnet 5.5. I love the like the documentation that comes out of that and uh at the end of every single flow a team of docs writers goes back to write documentation about what that phase implemented. And so I can just go to the GitHub repo and like look at the documentation as it's going. That's one one way to do it. Another way uh so you kind of asked about like is there a way to like pause it or stop it? Um and so there is uh another thing that we'll talk about later is um interrupting a session so you can interrupt the flow and say like actually like you're going the wrong direction. Please stop right here. Um you can also prompt Claude to give you reports or like to update a certain report as you go or update it you know like a typically it'll like um give you these marked out files that you can open.

嘉宾：是的，产品界面有一件挺酷的事：不管你用不用 flows，如果你让 Claude 把内容写到它的 outputs 目录里，就会得到一组可下载的文件。这些文件非常持久，我们还有另一个叫 V1 files 的 API 接口。或者你可能会像 CJ 一直在做的那样，把提交推送到 GitHub。所以这些运行过程中会产生很多输出，你可以自己检查。对，所以问题是：如果你在解决某个疯狂宏大的问题时进行到一半，突然从外部发现了一些信息，可能会改变整个运行的走向，你会想就这么烧掉两天的 token，然后从头开始吗？Guest: Yeah, and one thing that's pretty cool about the um the the product surface is that you like no matter what type whether you're using flows or not um if if you have Claude write to the out its outputs directory you have like a set of downloadable files that are like totally uh durable with that another API surface we have called V1 files. And so or or maybe you're like writing commits to GitHub like um CJ's has been and so there's just like a lot of like outputs um as part of these runs that uh you can inspect yourself. Yeah, so the question is like halfway through some crazy ambitious problem you find some information like externally or something that can potentially change the trajectory of the the overall run. Uh do you just want to burn two days of tokens and like start from scratch?

嗯，所以现在的做法是中断会话，然后补充更多信息。比如说，引理 X 实际上可能对你很有用，Claude。看看这个。嗯，而 Claude，也就是你正在交互的主要代理，掌握着流程进行到哪一步的完整状态，因此它可以检查自己基本上是从哪里停下来的，然后启动一个新流程，同时考虑这些信息以及之前的历史记录。嗯，目前还不支持的是，你不能拿第一个流程，也就是第一个计划，在中途插入单个步骤，然后重新运行后续步骤之类的。不过，对，这就是你接收更多信息的方式。Um so the way this would work today is you would interrupt the session and like provide that more information, you know, lemma X is actually probably probably useful for you Claude. Take a look at that. Um and Claude like the the primary agent that you're interacting with has this like full state of where the flow is in its process and it can inspect where where it basically left off in order to uh kick off essentially a new flow that takes into account that information as well as the past history. Um what's not supported today is like you can't like take the first flow, like the first plan, and like inject a single step like halfway through um and like rerun later steps or anything like that. Um but yeah, that's like how you would ingest more information.

主持人：呃，问题大概是这样的：作为一名游戏开发者，假设你想通过类似 flows 这样的方式添加游戏内遥测。嗯，我想确认一下……如果我没答到你的问题上，请告诉我，因为我可能说偏了。嗯，我在这里想到的遥测，是指智能体运行时正在发生的事情，对吧？你可以把智能体放进任何你想要的产品体验中，无论是游戏、网站，还是其他什么。嗯，但遥测更关注的是当下正在发生什么。我是否需要根据会话正在进行的情况采取行动？我说的方向对吗？好，太好了。嗯，作为我们 API 接口的一部分，我们提供了一个极其丰富、持久的事件流，你可以实时接收所有正在发生的事件，也可以通过一个静态 API 端点来列出这些事件，然后我们……嗯，我想想。Host: Uh so the question is like as a game developer, let's say you want to put in like in-game telemetry um via something like flows. Um I I I want to make sure that I I'm Tell me if I I don't answer your question because I might I might be going in the wrong direction. Um so when I think of telemetry in this case, I think of like information about what's happening as the agent's operating, right? So you could you could drop an agent into whatever product experience you want, whether it's a game, whether it's you know, your website, whatever. Um but the the telemetry aspect is more like what's happening right now. Do I need to take actions based on like what the the the session is doing? Am I going in the right direction here? Okay, sweet. Um so we have as part of our API surface, we have this extremely rich, durable event stream where you can stream in all the events as they're happening in real time or you can like list them via a static API um endpoint and then we um Yeah, I'm trying to think.

我不知道我们是否已经I don't know if we've

主持人：它是多模态的吗？Host: Is it multi-modal?

嘉宾：所以，如果你想在 Claude 托管智能体里，比如说把游戏截图放进去，然后让它采取一些行动，我觉得以今天的情况来看，用户体验会相当糟糕，因为这会花太长时间，具体要看你想在游戏里做什么。不过我想说，Claude 托管智能体确实支持图像输入，而且 Claude 检查图像的能力真的很强，或者说正在变得越来越强。所以完全有可能，你只要把截图放进去，Claude 就能根据你想做的事采取相应行动。现在还不会有实时的感觉，但，对。哦，对。对，对，完全同意。这其实是个很好的观点，比如用遥测数据来调试崩溃之类的问题。Guest: So like if you want to throw in like basically in Claude managed agents, if you want to take like a game screenshot and like throw it in and like take some action, um I think today that would be a pretty bad user experience because it would just take too long depending on what it is you're trying to build in your game. But I will say that Claude managed agents does support image inputs and Claude is like really good or getting really really good at inspecting images. And so you could there's totally a world where you like just drop in the screenshot and like Claude can uh take some action uh depending on whatever it is that you want to do. It's not going to feel like real time right now, but yeah. Oh, yeah. Yeah, yeah, totally. Yeah, that's actually a a great point, which is like maybe telemetry like debugging a crash or something.

你可以把游戏里正在发生的事情截个图，然后把它丢给一个托管智能体。这完全完全可行。所以，问题是关于分享这些东西，对吧？比如你想分享这些资源。也许你是一个组织，有一组模板化的智能体，对吧？它们对你们很有用，你也想确保不只是你一个人能访问它们。目前这件事的实现方式是，首先，我们有一个 Terraform provider，你可以用 Terraform 来帮你创建智能体。另外，我们还在控制台里提供了很多类似导出功能的东西，可以让你导出智能体，再把它们放到其他地方；但目前还没有那种一键分享按钮，让你把智能体分享给其他人。基本上，我们只是努力降低各种门槛，让你可以获取智能体的主体，或者说，任何 API 资源。You like take a screenshot of what's happening in the game and like throw it into an managed agent. Totally totally feasible. So, the question is about sharing these things, right? Like you want to share these resources. Maybe you're like an organization that like you have like a set of templated agents, right? They're useful for you and you want to make sure it's not just you who has access to them. The way that this would work today, number one, we we have a Terraform provider that's like you can use Terraform to like create agents for you. But also we have a bunch of like export style functionality in in the console that lets you export agents and drop them into other places, but there's not like a one-click share share button where you like share an agent with other people. Basically, we've just tried to break down the barriers to make like taking the body of an agent or you know, whatever API resource.

我们尽量让你能轻松地把它拿去上传到别的地方。不过目前 API 内还没有原生的分享机制。We've tried to make it as easy as possible to take that and just like upload it somewhere else. But there's not like a first-class like sharing mechanism within the API right now.

主持人：我想说，最常见的做法可能是为这个工具构建一个产品界面，再把它提供给客户。比如，你可以自己搭建一个网站或应用，客户访问后，就能通过你构建的界面与它交互；这个界面只是在你托管的智能体之上加了一层。因此，很多 Claude 托管智能体的交互都是直接通过 API 进行的。许多客户会把这项功能直接集成到产品中，提供给自己的终端客户。比如，你在一款非常流行的项目管理工具中使用看板，把一张卡片从某一列拖到另一列，这可能就会触发一个 Claude 托管智能体会话，让它通过 API 启动一个流程，或开启某个会话，在后台执行一些操作。Host: I would say like the most common way that you might give this tool to a customer is to build a product interface for it. So, like you would build your own like website or something or app and they come to that and interact with it through something that you've built that is just like a layer on top of your managed agent. And so, like a lot of a lot of Claude managed agent interaction is just through the API directly. And many many customers are offering that to their end customers as like built like just built into the product. So, like if you're you're in very popular project management tool that has a Kanban board and you drag one card from like a column to to next column, that might trigger a Claude managed agent session to go like, you know, using the API, spin up a flow or spin up some session that does something um in the background.

所以，它不是那种可以直接使用、能在 Claude 各项服务之间共享的前端组件，而是你会集成到自己产品中的东西，嗯，更多是在幕后运行。对。And so, instead of having it like be a like front-end component that is directly available that you can kind of like share out across Claude services, it's something that you would build into your own product um that's kind of more behind the scenes. Yeah.

嘉宾：根据给他们的东西是什么。所以，问题是，有没有办法调整智能体的行为，让它根据你交给它的工作类型来决定，而不是在智能体的系统提示词里做设置？嗯，这个回答有点敷衍，不过我可能会直接在系统提示词里说明，比如，如果你希望智能体根据工作的规模来考虑问题——假设你大致知道工作的规模——那么智能体就应该根据规模采取不同类型的行动。没有什么可以调节的旋钮……好吧，其实我收回刚才的话。根据你具体想实现什么，我们可能有办法做到：我们有按会话设置的 effort 旋钮，也可以按会话覆盖某些字段。Guest: By the what which is being given to them. So, the question is like is there a way to tweak the behavior of agents rather than like in the system prompt of the agent instead to be keyed based on the type of work that you give the agent? Um a little bit of a cop-out answer, but I would probably just system prompt that based on the like like like if you want the agent to think about let's let's say that you know rough sizes of work, like the agent should like take different types of action based on that size of work. There's not like a knob Okay, well, actually I I take that back. One way that you might be able to do this um depending on what exactly you're trying to accomplish um is that we have effort knobs on the um uh on a per session basis. And we also have the ability to like override some subset of fields on a per session basis.

所以，如果你有一个任务和一个智能体，就可以设置这些 effort 参数，或者根据你想要完成的目标调整智能体的不同部分。我大概会这么做。对，所以问题是，Claude 到底是怎么知道的，对吧？Claude 怎么知道该以什么规模运行？嗯，我知道这可能不是最让人满意的开放式回答，但答案其实就是评估和更低的默认设置。比如说，如果你让它运行股票行情查询，我们不应该启动一千个智能体，让它们都消耗大量 token，对吧？但如果你问的是纳维-斯托克斯方程，对吧？那也许就得在流程上再加把劲，明确要求它启动相应级别的智能体，也就是相应级别的并行处理。对。And so, if you like given a task and an agent, you can like set those effort knobs or like change different parts of the agent based on what it is you're trying to accomplish. That's probably how I would do that. Yeah, so the question is how does Claude even know, right? Like how does Claude know what scale uh to to operate? Um my I know it's probably not the most like satisfying open answer, um but the answer really is like evals and lower defaults where you like if you're if you're asking like the stock ticker thing to run, like we shouldn't be spinning up a thousand agents that are all burning a ton of tokens, right? But if you're asking about Nav your Stokes, right? Like maybe you would have to be pushing a little bit harder on the flow and like asking it specifically to do that uh like like spin up that uh level of agent yeah that level of parallelization so yeah.

好吧，我觉得我们大概还是应该继续。我们还有一个。最后一个问题，然后是的，最后我们也可以再回答更多问题。明白了，所以问题是：基本上用 Markdown 文件来查看代理在做什么，同时大概也用来定义代理等等，这是不是正确的做法？我觉得这个问题很难一概而论。实际情况是，我们的 API 接口根据你想做的事情，提供了非常非常多的功能。举个例子，我们会聊到凭据保险库；如果你把 API 密钥放进一个 Markdown 文件里，然后直接交给 Claude，那可能不太好——抱歉，我应该说，这肯定不好。千万别这么做。所以这真的取决于你想通过 API 接口表达的内容有多复杂。All right, I think we probably should keep rolling. We got one more. One last question and then yeah we can also take more questions at the end too. Got you so the question is is using basically like markdown files to both like see what the agent is doing but also presumably like define the agent etc. The right way to do it. I I think there's a lot of It depends for lack of a better answer. The the reality is that our API surfaces have like a ton ton ton of functionality depending on what it is that you're trying to do. Like we'll talk about credential vaults in an example like if you throw in your API key to a markdown file that you just like give to Claude that's probably bad and sorry I should say that's definitely bad. Do not do that. And so it really just depends on like the complexity of what what you're trying to express with the API surface.

我想说，如果你只是想要一个基础的 Markdown，再配上一套基础的……比如说，基本上就是把它交给 Claude，让它处理一下，偶尔用一次，那 Claude manager agent 可能就不是合适的抽象方式。不过，Claude 在底层的很多工作方式，你知道，基本上都是文本，所以其实也没有差得太远。好，我们继续。I will say that if you're just trying to like literally have like a basic markdown with like a basic set of like like Claude basically take that and like do something with it on kind of a one-off basis then Claude manager agent probably isn't isn't the right abstraction for that but like a lot of what how Claude works under the hood with you know all this stuff is basically like text right so it's it's not too far off. Okay let's okay.

主持人：对，我再补充一点，然后我们就可以继续了。需要考虑的一点是，某些上下文对每个子代理有多相关。一种做法是逐步呈现这些上下文：用一个 Markdown 文件链接到其他 Markdown 文件。比如，如果你属于负责内核的工程团队，那么主文档可以列出各个独立文档的索引：内核在这里，汇编器在这里，编辑器在这里。这样，子代理就可以读取与自己相关的详细内容，也能帮我省下一些 token。嗯。Host: Yeah I'll just add one more thing and then we can keep going. So one like thing to consider is like how relevant certain context is to each sub agent and so one technique is to gradually reveal that context by having like a markdown file that links to other markdown files so like if you are part of the engineering team that's working on the kernel maybe you have like the main document will say okay here's the separate docs like an an index of like kernel is here and assembler is here and editor is here. And so like it can then read the details that are relevant to that sub agent and I can kind of like save some tokens. So.

主持人：可以开始了吗？Host: We good to roll?

嘉宾：我们开始吧。好。Guest: Let's do it. Yeah.

主持人：好吧，抱歉。我们得继续了。最后我们会留出时间提问。所以先把问题留着，呃，我猜结束后我们可能也会再待一会儿。我们甚至还没过来看看大家的控制台。我希望它能正常运行。呃，如果没成功，我们很乐意稍后帮你们跑通第一个流程。Host: All right. I'm sorry. We've got to keep moving. We'll we'll have questions at the end. So so save your questions and we'll uh we'll also probably hang out for a little bit after if I I had to guess. We didn't even come around and like look at people's consoles. I I I hope it worked. Uh if it didn't, we're happy to help you uh run your first flow uh later on.

嘉宾：嗯，好。之前我们聊过，或者说 Harrison 提到了 Bug Hunter；上一次 AIE 中浮现出的另一个常见主题，则是围绕循环展开的。我们该如何闭合这些循环，让所有这些系统能够更自主地运行、自己完成更多工作？Bug Hunter 是一种非常有效的工作流和模式，我们一直在使用。我想稍微聊聊它。它也可以作为模板使用，也能在 Cloud Code 中使用。Bug Hunter 背后的想法是，你可以提交一个 PR，然后让 Bug Hunter 工作流审查这个 PR，并找出可能存在的问题，比如性能问题，或者实际的故障等，从而让你作为人工审查者能够更快地审查 PR。所以，与其只让一个代理审查 PR，不如用 Bug Hunter 这种更有效的方式来审查 PR，因为它会使用一个……Guest: Um okay. So earlier we talked about or uh Harrison mentioned Bug Hunter and another common theme that fell out of the last AIE was around loops. How can we close these loops to make all of these systems kind of like work more on their own and run more autonomously? And Bug Hunter is a very very effective flow and a pattern that we have been using um and I I want to talk a little bit about it. It's also available as a template and also in cloud code. So the idea behind Bug Hunter is you can put up a PR and then you can have a Bug Hunter flow review that PR and uh surface problems that might be uh performance issues or uh you know, like actual like uh breakages um and uh allow you to more quickly review PRs as a human. And so uh instead of just having a single agent review a PR, Bug Hunter is a is is a much more effective way to go about reviewing PRs because it uses a

Bug Hunter：多智能体 PR 审查Bug Hunter: multi-agent PR review

流程是这样的：它会接收你的代码，把代码拆分成这些片段，再将每个片段分别交给一个独立的上下文。你可以把它想象成有一个代理，负责只审查其中一个代码片段，并开始记下“这可能是个问题，那可能是个问题，这也可能是个问题”。然后把这些内容交给第二阶段，让审查者针对这个代码片段，判断自己是否认同这些不同的问题。最后，再把所有结果汇总到一个阶段，由它挑选出可能确实存在、需要进一步核实的问题。比如说，如果你是 PR 的人工审查者，在查看代码时，往往会了解到代码库中其他一些部分的背景，而这些内容甚至没有在这个 PR 中提及。flow where it takes your code, splits it up into these slices, each of those slices is handed to an individual context. You have this like uh an agent that has uh the responsibility of reviewing only that slice of code and starting to like write down this might be a problem, this might be a problem, this might be a problem. Then handing that to a second phase where that reviewer can say like, "Okay, given this slice of code, I agree or disagree with these different problems, and then having that all be surfaced to one phase that can like pick candidates that might actually be real problems that we need to go out and verify. So, like if you were a human reviewer of a PR and you are looking through some code, you often will have context about different parts of the code base that aren't even mentioned as part of this PR.

因此，我们通常会相信人类继续审查即将进入代码库的代码。呃，不过，有了智能体主动向你指出这些问题，这个过程就会高效得多。人类可能会做的一件事是这样说：“我觉得这会有问题，因为我知道代码库的另一个角落里有这样一个问题。让我拉取这个 PR，确认它是否真的会造成影响，以及影响可能有多大。”接下来，你可以进入这个验证阶段：下载代码库，让智能体尝试复现这个 bug，或者验证它确实是个实际问题，然后将结果整理成报告。在这个例子里，我们只是说它是一份报告，但在实际使用中，这些内容会作为评论添加到你的 PR 上。Bug Hunter 会直接在相关代码行旁指出问题，并说：“好，你正在修改这个文件，而代码库里另一个相距甚远的地方会因此出现意料之外的问题。”And so, that's often why we trust humans to go and continue to review the code that's entering our code bases. Uh but this becomes much more effective because the agents can start to surface those problems to you. Um one of the things that you might do as a human is say like, "I think this is going to be a problem because of this this other like corner of the code base that I know about. Let me pull down this PR and verify that that's actually like going to be an issue and how big of an issue it might be." And so, you can run through this prover phase where you could download have an agent download the repo, try to reproduce the bug, or like verify that the bug is actually like a legit problem. And then have these surfaced in a report. In this case, we just say like, you know, it's it's a report, but in practice, these are added as comments to your your actual PR where Bug Hunter would raise an issue like inline and say, "Okay, you're making this change to this file and then way off on like this this distant corner of the code base, that's going to cause like this unexpected problem for you."

嗯，它可以给你优先级，你知道，比如这是一个非常严重的问题，或者没那么严重的问题。嗯，所以再说一次，这是一个例子：你可以根据自己的代码库以及项目构成（比如是单体仓库，还是有很多不同的仓库等等），把测试时的计算资源分配给问题的不同部分。嗯，Bug Hunter 是一个非常非常有效的工具，可以帮助你在 PR 流程中形成闭环。你可以用 Cloud Code 这样启动它：Cloud API、扩展规模、管理代理、接入 Bug Hunter。嗯，呃，对，这大致就是它的工作方式。还有一个非常有趣的工作流。嗯，有多少人关注了我们的 Twitter 账号 Cloud Devs？你们有关注 Cloud Devs 吗？Um and it can it can give you priorities, you know, this is like a very severe or like less severe thing. Um So, again, this is a one of the instances where you can like throw test time compute at different parts of the problem depending on your code base and the composition of like, you know, is it a mono repo or you have like lots of different repos, etc. Um Bug Hunter, really really effective tool for uh helping close the loop on um your PR process. This is how you would spin it up with Cloud Code. Cloud API scale manage agents on board Bug Hunter. Um and uh yeah, this is kind of like how how that works. Another really fun flow. Um how many people follow us on Twitter, Cloud Devs? Do you follow follow Cloud Devs on

解说视频工作室：制作视频的流程Explainer Studio: flows that make videos

Twitter？好。行。哦，你们人还不够多。还不够多。嗯，好吧。所以，呃，如果你在 Twitter 上关注 Cloud Devs，就会看到我们很多时候在发布新产品或新功能时，都会配上一段动画视频，介绍该产品或功能的工作原理。实际上，我们会用一个工作流来生成很多这样的视频。所以，呃，你可以——我可以直接启动一个工作流，说：“讲讲动态工作流是如何运作的。”然后它就可以依次执行其中的各个阶段，甚至可以作为子工作流来运行；或者，呃，不好意思，不是子工作流，而是单独的、彼此独立的工作流，如果我们想这么做的话。呃，它会有一个研究阶段，会去查找有关其工作原理的所有实际细节，然后进行规划、探索和事实核查。接着它会整理出一份简报。Twitter? Okay. All right. Oh, not enough of you. Not enough of you. Yeah, okay. So, uh if you follow Cloud Devs on Twitter, you'll see that a lot of times when we announce new products or new features, they come with like an animated video that explains how that product or feature works. So, we actually use a flow to generate many of those videos. And so, uh you can I can just say, you know, kick I can kick off a a flow and say, "Explain how dynamic workflows works." And then it can run through each of these phases or even uh as sub flows or like I'm sorry, not sub flows, just individual separate flows if we wanted. Uh so, it's going to have like a research phase. It's going to go out and find what are all the actual details about how this works. And plan and scout and fact-check. And then it will build a brief.

然后，它就可以把这份设计简报交给制作团队，由他们制作各个独立场景并为场景添加动画。实际操作中，我们还会经过另外几轮流程，处理音效、配音、脚本等各种其他工作。最后，你就能得到一段视频。 嗯，Explainer Studio 的整个流程用起来确实很有趣。作为一个稍微负责打理我们 Twitter 账号的人，我觉得它为我节省了大量时间。你知道，大概四年前，要制作我们发布在 Twitter 上的视频，我得花上大概 20 个小时来做动画、把视频制作得漂亮。而现在，只要给这个流程一个很简短的提示就行了。 你也可以在 Cloud Code 中使用搭载 Cloud API skill 的托管代理 Explainer Video Maker 来完成这件事。And then it can take that design brief and hand it off to production that will go and make individual scenes and animate those scenes. And then uh in practice, we actually have like another set of phases that will do like sound effects and voiceovers and scripting and like all this other stuff. And at the end, you get a video. Um so, Explainer Studio, this is like actually like a quite quite a fun flow to use. Um and uh as someone who helps a little bit with our Twitter profile, I think it like saves me a ton of time. I think like, you know, 4 years ago, making the videos that we publish uh on Twitter would have taken like, I don't know, 20 hours or something to animate and like make beautiful. And now it's just, you know, one very small prompt off to uh off to a flow. Um you too can do this in Cloud Code uh with the Cloud API skill managed agent on board Explainer Video Maker.

嗯，对，就问一下 DNS 是什么？DNS 到底是怎么工作的？我也不知道。我们看看。呃，它会生成一个流程。我们已经有，嗯，它生成了几个视频。那我们来看看。好吧，如果任务的环节太多，一个代理处理不过来怎么办？好吧，我们有 300 份合同。我们想让它逐一检查这些合同，找出一些信息。通过动态工作流，代理会编写工作流，然后找出这些控制权变更条款。后台服务器会运行这个工作流。这是……嗯，这个只是我的声音。对，因为我们……Um and yeah, just ask how What is DNS? How does DNS even work? I don't know. Let's see. Uh and it'll generate a flow. So, we've got Yeah, we've got a couple uh of videos that it made. So, let's let's see. All right. So, what if the task has too many pieces for one agent? All right. So, we have 300 contracts. We want it to go through all those contracts and find something out. So, with dynamic workflows, the agent is going to write the workflow. And then find these change of control clauses. So, the server in the background is going to run that workflow. This is well, this one is just my voice. Yeah, cuz we

主持人：是啊，呃……听起来很棒，对吧？并不是。结果返回后，它们会全部合并起来。这些就是我们刚才讨论过的不同事件。你可以看到，它的动画解释做得相当不错，说明了会话线程是如何工作的，以及如何通过 API 传递这个 multi-agent 参数来启用它。接着它还会告诉你：“哦，对了，设置预算可能很重要。如果预算发生变化，或者你提前结束、还没花完预算，也可以延长预算，等等。”嗯，是的。Host: were uh yeah. Sounds so good, right? No. The results come back. They're all combined. So, these are like the different events that we talked about. So, you can kind of like see it's actually like pretty well animated explaining, you know, how session threads work, how to enable it through the API by passing this multi-agent uh argument, and then telling you, "Oh, yeah, it's probably important to set up a budget. And if the budget changes, you can or like if if you if you if you uh end early without spending it, you can extend it, etc." Um yeah.

嘉宾：我该如何发布到 YouTube 或者Guest: How do I deploy to YouTube or

主持人：它不会部署到 YouTube，不会。呃，还有……Host: It does not deploy to YouTube, no. Uh and

嘉宾：嗯，这是个好问题。嗯，我们确实应该这么做。应该的。对。所以，如果你……呃，想在 YouTube 上关注一个不错的频道，可以关注 YouTube 上的 Claude AI 频道。我们在那里做过一些特别有趣的事，比如直播音乐。Claude FM 最近还直播了一场很有秋日氛围的音乐节目。谁想看看自行车是怎么造出来的？好。有多少人想看看马桶是怎么造出来的，或者马桶是怎么工作的？对，不是看它们怎么造出来的，是看它们怎么工作的。好吧。行，我们来看看。对，我觉得马桶赢了，不过优势不算特别大，只是……Guest: Um that's a good question. It's a um We should, though. We should. Yeah. Right. So, if you if you uh a good follow on YouTube is the Claude AI channel on YouTube where we have done some like really fun things where like it'll stream like music and uh Claude FM did like a very kind of like nice autumnal kind of like vibe music thing uh live streamed recently. Who wants to see how a bicycle is made? Okay. How many people want to see how a toilet is made or how a toilet works? Yeah, not not how they're made, how they work. So, okay. All right, let's see. Yeah, toilet I think was the winner just like not by a ton, but

主持人：是啊，我觉得[清了清嗓子]我们知道自行车是怎么运作的，所以，没错。而且我也很喜欢这些视频看起来像是打上了 Claude 的品牌标识。比如我们让它做，它就会给我们一个品牌组合，里面有一套颜色什么的，嗯，用作背景；你也可以给它自己的品牌风格。还有一个建议是用 Claude Design 给它设定你的品牌风格，然后一步步创建自己的设计系统。你可以打开 Claude AI / Design，也可以直接在 Claude 桌面应用里使用 Design，然后把你在 Claude Design 里设定的风格交给你的工作流，再说“照着我的品牌风格设置来做”之类的话。就连底部的字体都是 Anthropic 的字体，这点也挺酷的。Host: Yeah, I think [clears throat] we know how bicycles work, so Yeah. Also love that it's like Claude branded the videos. Like we ask them, it'll give us like a lockup that has, you know, sort of the the the same colors and everything Uh for the backgrounds and you can kind of like give it your own brand style. Another like plug would be to use Claude design to give it your brand style and go through and create your own like design system. So, you can go in like Claude AI {slash} design or use design directly from the the Claude uh desktop app and then given your Claude design style, you can uh give that to your flow and say like, you know, follow my brand uh my my brand settings and it even like the font at the bottom is even like the Anthropic fonts and everything, which is pretty cool.

所以，这真的是一个非常强大的工具，尤其是当你负责制作动画视频并将其发布到网上时。嗯，对。嗯，这是一个关于流程的另一个有趣例子，不过它有很多很多不同的使用场景：无论你是想解决某个……你知道的，极其重要的千年难题，还是想生成视频，或者想改进自己的股票选择，抑或是……你知道的，开发流程。[清嗓子] 用流程可以构建很多很多不同的东西。嗯，我们稍早前提到过，单个智能体在一些事情上会遇到困难。这是我们在动态工作流中学到的一大经验，我想简单谈谈这些困难，以及它们如何适用于…… sort ofSo, really really powerful tool, especially if you are responsible for animating videos to post on the internet uh and yeah. Uh another really fun example of flows, but there's lots and lots of different use cases, whether you're trying to solve some um you know, really important millennium problem or you're trying to generate video or you're trying to like improve your stock picks or uh you know, your dev flow. [clears throat] There's lots and lots of different things that you can build with flows. Um so, one thing that we talked about a little bit earlier was a a handful of things that single agents struggle with. And this is a big takeaway that we learned as part of dynamic workflows and I wanted to just talk a little bit about them and how they apply to uh sort of

懒惰、自我偏见与目标漂移Laziness, self-bias & goal drift

来自 Claude Code 的内容，以及它们如何适用于 Claude 托管智能体。其中一个是智能体惰性，另一个是自我偏好偏差，第三个是目标漂移。嗯，结果发现，如果我们的提示词是这样的：“我们现在遇到了故障。请去检查这 50 个不同的服务，告诉我……呃，告诉我这次故障的根本原因，但不要碰生产环境。也就是说，实际上不要在生产环境中更改任何东西。去调查一下是不是发生了故障？不要碰生产环境。”单个智能体可能会遇到这样的问题：它看到 35 个服务都正常后，就开始说服自己，觉得“哦，看来一切都挺好。应该没事。我其实不需要再检查剩下那些服务了。”所以，嗯，我们发现，智能体惰性可能是单个智能体会遇到的问题之一；看过很多服务都正常后，它可能就会觉得：“好了，我觉得没问题了。”from from Claude Code and how they apply to Claude managed agents. So, one is agentic laziness, another is self-preferential bias, and the third is goal drift. So, um it turns out that like if our prompt is something like, "We're having an outage right now. Please go look at these 50 different services and tell me if uh you know, tell me the root cause of uh of the outage, but don't touch production. Like don't actually like change anything in production. Go research is there an outage? Don't touch production." A single agent might have this problem where after it sees 35 services are healthy, then it starts to like convince itself that oh, it's like everything looks good. It's probably fine. Like I don't really actually need to look at the rest of them the services. So, um we have seen that agentic laziness can be like one of the things that a single agent will run into and after it's seen so many that are good, it can kind of just be like, "Okay, I think we're good.

“我就不看剩下的内容了，省点 token，告诉告诉我们的人工用户，一切都准备就绪。”协调器也会遇到类似的问题：可能有三个子代理向报告代理或协调器汇报，而协调器仍然得跟踪：“好，我们已经看到几个了？”以及所有服务是否都正常？这些子代理自己到了一定时候也可能变得懈怠。所以，嗯，和只有一个代理相比，实际上并没有太大改善。有了工作流之后，我们有这样一种方式——它不一定是确定性的，但我们可以拆分任务，并为每个服务安排一个独立的代理。这样一来，它们只负责各自范围很小的工作，我们就能得到好得多的答复，因为我们可以通过工作流，让任务扇出到全部 50 个服务，并为每个服务安排一个代理去检查该服务是否正常。Let me just not look at the rest and let's save some tokens and tell tell our human that like we're good to go." Coordinator runs into a sort of a similar problem where the you might have three sub-agents reporting back to the reporter or to to the coordinator and the coordinator still has to like keep track of like, "Okay, how many have we seen?" and are all of the services healthy? And those sub-agents might also themselves get lazy at a certain point. And so, um you don't really like improve too much over having the single agent. When we have flows, we have this sort of it's it's not necessarily deterministic, but we have the ability to split up and dedicate a an individual agent in context to each of those services. So, they're only responsible for their small scope of work and we can get a much better answer back because we can tell from the flow you know, fan out across all 50 services and dedicate an agent to each of those services to go and see if that service is healthy.

到最后，我们就能确定已经逐一检查了所有服务。所以，这有点像是把扇出和循环式智能结合起来。自我偏好偏差。嗯，如果你让一个智能体写一首诗，然后问它：“写得好吗？”它会回答：“好啊。”嗯，这是很常见的情况。如果你让一个智能体写一些代码，然后问它：“这段代码写得好吗？”它也会回答：“写得好。”这是因为通常是它自己完成了这项工作，而且它对自己的成果很满意。所以，如果你能把这两个环节分开，就能得到更好的结果。我们从顾问模式中了解到这一点：你可以让一个智能体负责生成，再让另一个智能体评判生成的结果。这有点像是利用了生成器和判别器之间的差距。And by the end, we know for sure that we went and checked every single service. So, this is like the fan out and loop intel done kind of like combined. Self-preferential bias. Um if you ask an agent to write a poem and then you ask the agent, "Is that good?" It's going to say, "Yeah." Um this is like a pretty common thing. If you ask an agent to write some code and then you say like, "Is the code good?" It's going to say, "Yeah." And it's because it like generally was the one who did the work and like is happy with the work that it did. And so, if you can uh split that up, you get better results. We know this from the advisor pattern where you might have like one agent that is the generator and one agent that is like judging the results of that. Like kind of uh taking advantage of that generator discriminator gap.

但是，当你遇到一个特别棘手的大问题时，还是会碰到这种……你会遇到这样的情况：你的智能体阵容未必能扩充到所需规模，除非设置这些彼此独立、各自分明的检查点。因此，借助工作流，我们可以把任务分散给许多不同的智能体，并进行对抗性审查；审查有自己隔离的上下文，完全不受之前内容的影响。But, when you have like a really massive problem, you're still going to like run into this prothis you'll you'll run into a situation where like you your roster is not going to like necessarily scale up to a point that you need without having like these separate and distinct kind of like checkpoints. And so, with a flow, we can spread out across lots of these different agents and have like adversarial review that has its own isolated context that isn't being like influenced at all by previous things

大规模智能体安全性Agent safety at scale

它检查过的内容。最后一点是目标漂移。还记得最初的提示大概是：“去扫描所有这些服务，告诉我我们的状态是否正常。”但我最后说的是：“不要碰生产环境。”对吧？所以，如果我们不断填满上下文窗口、不断压缩上下文，在检查所有这些服务的过程中，我们不仅可能会偷懒，倾向于采纳自己的答案，而且“不要碰生产环境”这条要求也可能在这个过程中丢失。因此，有一点很重要：随着上下文越来越长，我们可能会逐渐忘记自己的确切目标。而使用工作流时，我们可以在一开始设定一个非常明确的目标；由于我们有这些独立的上下文，就能确保这个目标以非常一致的方式传递给各个阶段的所有子代理。that it checked. The last one is goal drift. So, if you remember like the the initial prompt was something like, go out and scan across all these services and tell me if we're healthy, but the last thing that I said was like, don't touch prod. Right? And so, if we are filling the context window and compacting and we're filling the context window and we're compacting, we're going through all these services, not only might we be lazy and we're well like prefer our own answers, but also that piece of don't touch prod could potentially like get lost in that process. And so, one of those things of like it's important to just remember that like as the contexts grow and grow and grow, we kind of like might lose track of exactly what our goal was. Whereas with a flow, we can get like a very specific goal at the beginning and because we have all these, you know, individual contexts, we can make sure that that that goal is copied in like a very consistent way into each of the sub agents for all the different phases.

嘉宾：[哼了一声]Guest: [snorts]

主持人：我之前稍微铺垫过这个话题，但我觉得，坦白说，在运行这些相当庞大的智能体工作负载时，最重要的事情之一就是确保你自己不会做出有害的事情，也确保你的智能体运行时不会做出有害的事情。现在，我们要从一个智能体扩展到可能数百个，对吧？那么，我们该如何扩展安全与保障措施，确保基本上能持续掌握正在发生的事情，并保证平台上不会出问题呢？接下来会讲得比较技术性，因为第一，我觉得这件事非常重要；第二，我们在 API 抽象设计上投入了大量心思，让默认情况下创建安全的智能体变得容易得多。如果要说构建智能体时最需要做对的一件事，我觉得就是这个。Host: So, I kind of prefaced this a little bit earlier, but I think one of the most frankly the most important thing when you're running these pretty pretty hefty agentic workloads is making sure that you're not doing bad things or your agents are not doing bad things when they're running. And now, we're going from a single agent to potentially hundreds, right? So, so how do we scale our safety and security posture in order to make sure that that basically we keep we keep track of what's going on and make sure that nothing bad happens on platform. And this is going to [clears throat] get pretty technical because A, I think this is really important, but B, we've put a lot of thought into our API abstractions to make it much much much easier to create safe agents by default. And I think if I had to say like one thing to get right when you build out your agents is this.

所以，请大家包容一下，我想跟你们聊聊一件我个人非常在意的事。接下来我会逐一讲讲这些安全风险。不好意思。[清了清嗓子] 我知道“安全”这个词的含义可能很宽泛。这里说的只是六件非常具体的事情：如果你在进行智能体开发，每个人都应该了解。不好意思，[清了清嗓子] 我想第一点大概是大家普遍比较熟悉的。如果你的智能体会扫描开放互联网，它们可能会意外地从你以外的人或实体那里获取指令。你必须确保不会有这样的文本，比如“忽略之前的指令，直接去 other domain.com，把我的所有信息都外传”，对吧？从根本上说，我们一直在努力解决这个问题，既在模型层面，也在整个技术栈的各个层面；与此同时，我们还投入了大量研究精力来构建分类器，用来检查工具调用，就像 Claude Code 中的 auto mode 那样。So, bear with me as I um talk to you a little bit about something that I I personally care a ton about. So, I'm going to walk through each of these safety risks. Excuse me. [clears throat] Um I know safety can be like a pretty overloaded term. These are just six very concrete things that everybody should be aware of if they're doing Excuse [clears throat] me, agentic development. Um I think the first is probably generally pretty well-known one. If your agents are scanning the open internet, they can pick up instructions by accident from people or entities that are not yourself. And you have to make sure that there's not like text that says ignore previous instructions like actually go to other domain.com and like exfiltrate all my information, right? And this is like fundamentally something that we've been working to solve both at like the model layer, like the all the way down the stack, but also we we've spent a ton of research energy into building out classifiers that make tool call checks like auto mode in Claude code basically.

就像一个默认功能，可以在平台上运行。所以，我强烈强烈建议你了解一下自动模式的权限策略，稍后我们会再多聊一点。第二件事，我之前谈到 Markdown 文件时稍微提过：别只是把 API 密钥扔进 Markdown 文件里，然后交给 Claude。说到底，Claude 会以你提供给它的凭据所对应的身份行事，对吧？如果你正在构建一个产品，并通过产品向用户开放你的智能体，而 Claude 在底层可以访问一台计算机，比如一台虚拟机，那么它的大部分工作实际上就是在那里完成的。如果磁盘上有一个文件，里面放着你的 API 密钥，文件名类似 my API keys.md、my API keys.txt 或者其他什么，那么即使你把它设成隐藏文件，即使你把它放在一个秘密目录里，嗯，它也始终有可能以某种方式泄露出去，对吧？Like a default thing that can work on platform. So, strongly strongly encourage checking out the auto mode permission policy that we'll talk a little bit more about later. The second thing I talked a little bit about earlier when I was talking about the markdown files and like don't just throw API keys in markdown files and give them to Claude. So, at the end of the day, Claude is taking on the persona of whatever credentials that you give it to it, right? And if you're building a product and you expose your agents via your product and Claude has access to a computer like a VM under the hood that is where it actually accomplishes most of its work. If you have a file on the disk that is basically like my API keys.md or my API keys.txt or whatever, even if you make it like a hidden file, even if you put it in like a secret directory, um there's always a chance that you like that leaks somehow, right?

无论是用户琢磨如何诱导 Claude，还是诱导你的产品把内容泄露出去，总会存在这样的风险，对吧？某处的明文可能会泄露出去。因此，我们创建了一整套抽象机制，确保凭证永远不必接触沙箱。这种解耦至关重要，我稍后会详细讲到。另一个，也就是第三个问题，是数据外泄。如果你的智能体可以访问敏感信息，你就要确保它不会把这些信息分享给你以外的第三方。此外，你还要确保智能体不会超出你要求它执行的范围。我之前稍微谈到过预算。平台上还可以设置其他限制。Whether it's a user figuring out a way to prompt Claude or prompt your product in order to get it out. Um there there's just always a risk, right, that that plaintext somewhere um could could uh make it out. And so we we created an entire suite of abstractions to ensure, basically, that credentials never have to touch the sandbox. And this is an incredibly important decoupling, and I'm going to talk about it extensively later. The other or the third one is data exfiltration. If your agent has access to sensitive information, you want to make sure that it's not sharing that sensitive information with third parties who are not yourself. Additionally, you want to make sure your agent is not overreaching what you've asked it to do. I talked a little bit about budgets earlier. There's other limits that you can set on the platform.

此外，你还要通过为智能体制定权限策略，确保它不会超出自身的能力范围运行，这样只有你授予智能体访问权限的工具才能实际运行。最后但同样重要的是，能够了解智能体正在做什么，可以说是基本要求。因此，我们提供了一个持久化事件流，几乎完整记录了智能体在整个生命周期中所做的一切。所以，作为开发者，你的工作有三项。Um on top of that, you want to make sure your agents are not operating outside of like their uh their wheelhouse by establishing permission policies on the agent so that only tools that you give the agent access to can actually run. And then last but not least, like it's pretty table stakes to have some visibility into what the agent is doing. And so we have a durable event stream that has essentially the full trajectory of everything uh your agent is doing across the entirety of its life cycle. So, your job as a developer is three

声明、包含、检测Declare, contain, detect

不同的事情。我们喜欢这个很棒、很 catchy 的说法：declare、contain、detect（声明、限制、检测）。嗯，但这就是实际情况，对吧？你必须声明你的代理是什么、它拥有哪些权限、能做什么、你希望它采用什么 persona、具备哪些技能、工具等等。嗯，至于 contain 这一部分，实际上说的就是代理在物理层面受到的限制。那么，如何限制代理，才能确保即便在最坏的情况下，它在开放互联网中可能造成的伤害也微乎其微？这就是为什么我们在环境策略中提供了非常明确的配置，你可以指定代理只能访问这些域名，或者只能进行网页搜索、不能抓取网页，反之亦然，具体取决于你的需求。最后但同样重要的是，我一直在说的持久化事件流，它是让你实时检测正在发生的事情的一种非常关键的方式。如果代理正在做某些你不满意的事——具体什么算满意，取决于你自己的定义——你随时都可以实时监听这些事件流，并立即中断任何偏离你预期方向的会话。different things. We like this nice like very catchy declare, contain, detect. Um but that's that's the reality of it, right? You have to declare what your agent is, the permissions it has, what it can do, what you what persona it should take on, skills, tools, etc. Um and then like the contain piece is literally like the the physical limitations of the agent. So, how do you contain the agent so that even in the absolute worst of case, the amount of harm that it could do on the open internet is is minimal. And so this is where we have very explicit configurations in your environment policy where you can say you can only access these domains or you can only access a web search but not web fetch or vice versa depending on on on your needs. And last but not least, that durable event stream that I keep talking about is like a very crucial way for you to to detect things that are happening in real time and if if the agent is doing something that you aren't are not happy with depending on whatever like your definition of happy is, and you can always listen to these event stream in real time and immediately interrupt any sessions that are going in a different direction than you'd like.

嗯，我想想。不错。那么我们当初推出 Claude 托管代理是在……Um, let's see. Cool. So when we shipped Claude managed agents back in

让大脑与双手分离Separating the brain from the hands

到了四月，几乎出现了这样一个梗：Anthropic 已经解决了这个争论，对吧？构建智能体基础设施时，大脑和双手应该分开。玩笑归玩笑，我们认为这是基础设施中非常非常关键的一个组成部分，它让安全性能够自然而然地成为默认选项。这里的想法是，大脑——也就是在通用智能体框架中实际运行推理的循环，以及会话状态，比如每个智能体和每个会话整个生命周期中发生过的事情——与我们所说的 Claude 的双手完全分开。双手是智能体实际能够在现实世界中采取行动的地方。你可以把双手想成是在 Claude 有权访问的计算机上运行 grep。April, there was almost like this meme that Anthropic like settled the debate, right? The the brain and hands should should be separate when you're building out an agentic infrastructure. And like jokes aside, we think this is a very very critical component of our infrastructure that makes safety kind of just default possible. The idea here is that the brain, which is like the loop that's actually running inference in the just general agentic harness, as well as the session state, so like what's happened across all of the life cycle of every agent and a session is fully separate from what we call the hands of Claude, which is the place that the agent is actually able to take action in the real world. And you can think of hands as like running a grep on a computer that Claude has access to.

你可以把它想象成一个定制工具，或者说，一个可能运行在你那台我们之前提到过的 Mac OS 设备上的定制工具。总体来说，应该绝不能让“手”获得访问权限，并以任何可能被视为有害的方式完全操控“大脑”。这正是我们如何将所有需要由“大脑”和“执行框架”访问的机密信息，与“手”本身隔离开来的关键。我们确保凭证注入有效的方式也很有意思。我记得的话，稍后会再详细谈谈；不过，核心想法是，任何能够访问机密信息的容器里，那些机密信息都是完全伪造的，对吧？也就是说，如果你试着用那些值去调用 my API.com，它们实际上并不能用。You can think of it like a custom tool that or rather a custom tool that might run on a you know, your Mac OS machine that we we talked about earlier. Broadly speaking, there should be like no way for you to given a hand access and totally steer the brains in any way that could be considered harmful. And this is like the key thing of like how we separate all of the secret information that can that needs to be accessible in the brain and the harness from the hands themselves. And so, the way um there's this really interesting way that we have made sure that uh credential injection works. And I think I talk about it on a little bit more detail later if I remember correctly, but um the idea is that every container that has access to any secret, you just totally fake the secrets, right? Like those values don't actually work if you try to call like my API.com with that secret.

只有在向外发送时，也就是右侧这个出口代理层，才会把这些凭据替换成实际凭据。所以，从技术上讲，根本没法提示 Claude，或者让它访问沙箱，再从沙箱里，或者从 Claude 的大脑里拿到这些凭据。这就彻底将智能体相关的部分与实际凭据本身隔离开了。 我之前好像已经先说明过这一点了，但基本上，默认情况下，你不应该因为使用 Cloud Manager 智能体或自行构建智能体工作负载，就不小心把自己的凭据泄露出去。所以，务必、务必、务必认真对待这一点。我觉得这非常非常重要。And only on the way out um at this egress proxy layer here on the right side, do those credentials actually get substituted for what like the actual credential is. So, there's no way like physically to prompt Claude or to um like give an access to a sandbox, get that credential off of the sandbox or off of off of Claude Claude's brain. Um so, this is is like totally decouples the the agentic stuff from the um the actual like uh credentials themselves. So, I yeah, I kind of preface this already, but basically like you by default uh like, you know, you should not be able to foot gun yourself um in using Cloud Manager agents or building any agentic workloads uh yourselves um into leaking your credentials. So, like definitely definitely definitely take this seriously. I think this is very very crucial.

嗯，那么我把话筒交给 CJ，请他来谈谈……Um So, I'm going to hand it over to CJ to talk about

嘉宾：是的，我们会逐一介绍这些不同的功能，以及如何通过 API 控制它们。再说一次，确保限制 Cloud Manager 智能体的一种重要方式——我们已经谈过几次了——就是设置预算。比如，我们可以说：“好，运行这个流程。预算是 1.25 美元。”如果花费达到 1.25 美元时流程还没完成，你就可以决定是否提高预算，让它继续运行。它不会重新做已经完成的工作。或者你也可以直接说：“好，这就是我们为这个特定会话设定的预算。”每次启动会话时，无论是不是流程，都可以使用这个预算。这与 Claude 托管智能体的另一个功能 Outcomes 配合得非常好。Guest: Yeah, so the uh we'll go through each of these different things and how you can control it from uh from uh the API. So, again, one of the important ways that you can make sure to contain uh Cloud Manager agents, and we've talked about this a few times, is setting a budget. And so, we can say, "Okay, run this flow. You have a $1.25." Uh and then if it hits $1.25 and the flow is not finished, then you can decide whether or not you want to up the budget and let it keep running. And it won't have have redo any of the work that it had already done, um or you can just say like, "Okay, that that's what we had budgeted for that specific session." Um so, every time that you kick off any session, whether it's flows or not, you can use this budget. And this pairs really nicely with um this other part of Claude managed agents called Outcomes.

因此，这两者的结合确实非常巧妙，因为现在你可以这样说：“我想要这个结果。这是我希望实现的成果和预期的评分标准，以及预算。”然后，在这个预算范围内完成这个目标。嗯，预算的作用同样是确保我们能控制开销，避免这些智能体意外地不断消耗 token。下面的第二个代码示例展示了如何通过 API 在流程进行到一半时中断它。你只需传入这个事件 `user.interrupt`，就能停止流程，然后可以重新向 Claude 发出提示，或者采取任何你需要的操作来继续处理。所以，这是两种不同的功能：一种是预算，另一种是在正在运行的会话中途进行中断。And so, the combination of these two things is really beautiful because now you can say, "I would like this result. Here is the outcome that I want and the rubric that I expect, and here is a budget." So, go accomplish this outcome given this budget. Um and the budget, again, is this way to just uh make sure that we're containing and not letting these agents just like uh continue to spend tokens um unexpectedly. The second uh code example there below is how you can interrupt a flow in the middle of the process from the API. So, you can just pass this event, user.interrupt, and this will stop the flow, uh and you can like then reprompt Claude or um do whatever you need to do to continue on. So, these are two different things. So, one is the budget, the other was you can interrupt in the middle of uh of a session that's running.

所以，测试这一点的一种方法是，拿你今天运行的流程，设置一个低于预期运行成本的预算，比如一美元左右，然后运行流程，再提高预算，观察它是否继续运行。这样你就可以稍微尝试一下。嗯，我们还提到了第二种情况：让一个审核代理回顾你所有的会话，了解每个会话中发生了什么，并查看调用了哪些工具，以及这些会话中是否发生了任何意料之外的事情。这是另一种使用场景，你可以启动一个 Claude 托管代理，并在提示中告诉它：“这是我希望你查找的事情。”And so, one thing to kind of like test this out is I would take, you know, take the flow that you ran today, and then set a budget that's lower than what you expect it to go over, maybe like, you know, uh a dollar or something, and then run the flow, increase the budget, watch it continue, so you can kind of like experiment with that a little bit. Um the second one we talked about the um this ability to have a reviewer agent go back through all of your sessions and understand what happened as part of that session and getting visibility into what were the tools that were called and did anything unexpected happen in any of these sessions? This is another use case where you could spin up a Claude managed agent with a prompt that says, "Here's the things that I want you to look for that happened."

而且，因为我们有这份持久化日志，也记录了发生过的每一件事，你可以通过 API 拉取这份记录，并把它交给一个托管代理，然后说：“审查这份完整记录，以及过去一周左右的所有记录，告诉我我的代理有没有做过什么出乎意料的事、访问过任何它们不该访问的内容，或者调用过什么有点奇怪或出乎意料的工具。”所以，嗯，你还可以做另一件挺实用的事：构建一个 Claude 托管代理，把今天的会话交给它，然后问它：“这些会话里有没有发生什么奇怪的事？”最后，我觉得这其实是最简单、最容易入手的做法——如果你现在已经集成了任何 Claude 托管代理，那就是把权限策略从“每次都询问”或类似模式改为开启自动模式。And because we have that durable log and we have every single event that happened, you can pull that transcript with the API and pass it to a managed agent and say like, "Review this entire transcript and all of these past transcripts over the last week or whatever and tell me if my agents did anything that was unexpected or accessed anything that was they weren't supposed to or called any tools that were kind of like odd or unexpected." And so, um this is another um kind of like a handy thing that you could do is build a Claude managed agent and hand it the sessions from today and just ask it like, "Was there anything odd that happened in these sessions?" The last and I think this is actually like the easiest lowest-hanging fruit uh if you have any Claude managed agents integrations today and that is uh to take your permission policy from either always ask or like uh and turn on auto mode.

开启自动模式。这真的超级超级重要，而且也是非常非常基础、很容易做到的事情。嗯，我们做了不少研究，也有一些公开研究表明，呃，你知道，当你的笔记本电脑开着，屏幕稍微掀开一点，你人在公园里，Claude 一直问你：“你批准吗？”大家就会说：“嗯嗯嗯嗯嗯。好好好好好。批准，批准，批准。”你知道，作为人类，我们有 97% 的概率会点击批准，所以，嗯，我们确实很容易就被带得只会说：“好，好，好。”但 Claude 要耐心得多；使用自动模式时，它可以判断某一次工具调用是否安全，还是会采取破坏性操作，以及它到底应不应该真的去做这件事。Turn on auto mode. This is like super super important and also very very low level or like low-hanging fruit. Um we did uh a bunch of research and there's some like public research about how uh you know, when your laptop was open and cracked a little bit and you're at the park and Claude keeps asking you like, "Do you approve?" And everyone's like, "Two two two two two. Yes, yes, yes, yes, yes. Approve, approve, approve." You know, like 97% of the time we're going to click approve as humans and so, um we are yeah, we easily get like lulled into just being like, "Yes, yes, yes." But Claude is like much more patient and using auto mode, it can like determine whether or not a single tool call is safe or it's going to take some destructive action and whether or not like it should actually like do that thing.

所以，自动模式下你只需要从“始终询问”切换到自动模式，或者从你当前的权限策略切换到自动模式，这算是最简单、最容易上手的一件事。所以。So, auto mode all you have to do is switch from always ask to auto mode or whatever your permission policy is today to auto mode and it's like the easiest lowest-hanging uh fruit thing to do. So.

凭据保险库与工具安全加固Credential vaults & locking down tools

主持人：太棒了。所以我想稍微详细谈谈前面提到的凭据库，以及为什么我们认为凭据管理是你用来运行智能体的平台中至关重要的一部分。在这个例子中，我们通过环境变量使用 Twilio API。天真的做法是，我们把容器，或者更准确地说，把会话拿来，然后上传一个包含 `.env` 文件的文件，里面有 Twilio API 密钥，这样智能体就能读取它。然后，如果智能体想要采取某种操作，比如运行 Python 脚本，就可以从那个容器里实际调用 Twilio API。不过，问题在于，这又会涉及我刚才聊了一会儿的那些明文内容。所以，你可以使用凭据和凭据库来管理这些机密，而不是像默认做法那样，直接把它们放进容器里。Host: Sweet. And so, I want to go into detail a little bit about the vaults thing that I alluded to earlier um and why we think credential management is such a crucial part of any platform that uh you run your agents on top of. So, in this example, we are using the Twilio API via an environment variable. So, in the naive situation here, we'd like take our container or our rather our uh session, maybe like upload a file that has this like.env in it that includes this Twilio API key so that our agents could read it, and then actually call the Twilio API from that container if it was trying to take some some action in let's say a Python script. Um but the problem with that again is like all the plain text stuff I've I've been talking about for for a minute now. Um and so instead of like that default just like getting thrown out of the container, um you can use credentials and credential vaults in order to manage these secrets.

嗯，所以在这个例子中，我们使用 Vault 凭据来创建一个……嗯，用来将我们的 Twilio API 密钥存储为环境变量。这个注入位置的设置基本上是针对所有从容器发出的网络流量：我希望密钥出现在哪里？我希望它在发出请求时作为一个……嗯，Bearer 令牌出现吗？还是说我希望它直接出现在请求正文中？嗯……你可以根据这个字段进行配置。另外，这里的网络哈希和这个允许的主机设置，基本上是说：如果我们要向域名 X 发起请求，嗯……也就是说，或者说，如果给定的是我们的凭据 X，我们应该在发出请求时将哪个域名替换进去？因为如果我们只是天真地猜出这些密钥，嗯，或者这些存在于容器中的占位密钥，我们可不希望我们的 Twilio API 密钥跑到……我不知道，比如 mydomain.com。Um so in this case we're using a vault credential in order to create a uh in order to store our Twilio API key as an environment variable, and this injection location thing is basically for all network egress on the way out of the container, where do I expect the secret to show up? Do I expect it to be in like a a bearer off uh token on the way out? Or do I expect it to be like literally in the request body? Um uh you can configure um based on this field. Then additionally, the networking hash here with this allowed hosting is basically where given a call to domain X, uh like what what where or sorry, given our credential X, what domain should we substitute on the way out? Because if we just naively were able to like guess these secrets um and or or these like placeholder secrets that exist on the container, we don't we don't want like our Twilio API key to ever go out to I don't know, like mydomain.com.

所以，安全地处理这些内容真的非常重要。创建好凭据并将其存入密钥库后，只需在创建会话时传入密钥库即可。平台会根据密钥库中的内容、其中存在哪些密钥，以及出口层中有哪些占位密钥，智能地判断该如何注入正确的值。 我们还讨论过，应该尽可能缩小智能体可用的工具集。比如，我们的智能体工具集中有两个我认为特别有用的工具：web fetch 和 web search。你可能希望默认禁用它们，因为你不希望智能体在未经你明确许可的情况下访问互联网。这种情况下，可以将你想禁用的任何工具类型的 enabled 字段设为 false。Um so really really important to um uh handle this stuff safely. And then once you've created your credentials uh inside of your vault, then you just pass in the vault and when you create your session and the the platform will be smart enough to know like based on what's in the vault and what secrets exist and at the egress layer what um like placeholder secrets exist, it can like figure out what uh how to inject in the correct values. We also talked about giving agents a the smallest tool set possible. Um so one of the like most um, I think interesting tools that we have in our agent toolset is web fetch and web search. And maybe you want to disable those by default because you don't want your agent to have any access to the internet um, without you very explicitly allowing it. In that case, you can set the enabled field to false for any of the the the types of tools that you want to disable.

嗯，下面是一些关于智能体工具的基本示例。MCP 工具很有意思，因为你事先并不知道 MCP 服务器会返回哪些 MCP 工具。比如说，你要调用 GitHub MCP。是的，我们信任 GitHub。不过，万一出了什么问题，他们的系统遭到入侵，攻击者往里面添加了一个类似数据窃取工具的东西呢？你大概不会想让智能体调用它。因此，你可以针对每个工具，按 MCP 工具集分别启用工具。此外，你还可以通过环境配置，限制智能体能够访问的网络范围。你可以再次把环境理解为 Claude 操作的那台计算机。Um, so, here's just some basic examples for agent tools. MCP tools are are a really interesting one because um, you don't know the MCP tools that are going to come back from an MCP server up front. So, like let's say that you're calling out to the GitHub MCP. Like, yeah, we we trust GitHub. Um, but like what if something happened and they were compromised and an attacker added like, you know, my exfiltration tool or something to it. Um, you probably don't want that to be picked up by your agent. And so, you you on a tool-by-tool basis can can enable um, uh, per per MCP toolset. So, um, additionally, you're able to limit the network access that your agent has access to um, like physically using the environment configuration. Um, so, you can think of environments again as like the computer that Claude is operating on.

因此，如果我只想让 api.twilio.com 可以访问，就可以直接在环境配置中明确写出来。这样一来，我的 API 密钥就被隐藏起来，不会遭到潜在的盗窃。而且我可以实际调用 api.twilio.com。如果我尝试调用其他网站，比如 website.com，那么在出站流量产生时，就会在防火墙层面直接失败。And as such, if I only wanted api.twilio.com to be accessible, then I can like literally literally say that in the environment configuration. And so, um, now I have like my API key on the like like hidden away from um, any potential like theft. And I can call api.twilio.com physically. If I tried to call, you know, some other website.com, it would like physically fail at the uh, firewall layer um, on as uh, like like when egress tries to occur.

总结与问答，第二轮Wrap-up + Q&A, round 2

嘉宾：好吧。总之，回到最初的话题，我们认为，很多时候，我们都在试图重新发明那些组织经验的轮子。你知道，我们在那个……七十年代学到过如何组建规模庞大的人类团队，让他们去完成这些工程壮举。我们非常期待让大家使用 flows，也很期待看看大家会用 flows 构建出什么。所以，再说一次，希望大家能借助这些工具，去完成一些真正非常宏大的项目。从模板开始，把它塑造成属于自己的样子。告诉我们你们的发现。是的，我们很乐意再回答一些问题。我想我们差不多要结束了，不过我们也会在会场外的走廊交流区待一会儿。当然，大家有任何问题，我们都很乐意解答。Guest: All right. So, just kind of bringing it back to the beginning, uh, we think that we're in many cases trying to like reinvent the wheel of these organizational learnings that we had, you know, in the the and '70s about how to build really massive teams of humans to go out and achieve these like engineering feats. And so, we're really excited to give you access to flows, and we're excited to see what you'll build with flows. So, again, hopefully you can take these and go achieve some really really ambitious projects. So, again, start from a template, shape it shape it to be your own. Tell us what you find, and yeah, we would love to take some more questions. I think we're getting close to the end here, but we can also we'll also hang out outside in the hallway track. Absolutely happy to answer any questions.

嗯，我看到这边有一个问题。是的，问题大概是，你试着给智能体布置一项任务，但它返回的结果不太容易验证。比如说视频制作，这在很大程度上取决于你的个人审美。于是你给它一些反馈，那么在工作流中，该如何反复迭代呢？我想，可以先从 Claude Code 入手，先对某个流程进行迭代。你可以在 Claude Code 中使用动态工作流，也可以直接尝试走一遍制作第一个视频的流程，然后让 Claude 把你的操作编码成一项技能，之后便可以在 Claude 托管的智能体中将它用作工作流的一部分。我觉得，通常一开始需要经过相当多的迭代，才能弄清楚有哪些不同阶段，以及这个流程实际上该如何运作。So, I see one over here, yeah. Yeah, so the question is like you try to you know, give an agent a task, and the thing that it gives you back is like not very verifiable. For example, with the video, that's like very much, you know, your human taste. So, then you give it some feedback, and how do you sort of like iterate on that within flows? I would say one place to start is to start with Claude code and iterating on some process first. So, you can either use dynamic workflows in Claude code or directly try to work through the process of generating your first video, and then have Claude encode like what you did into a skill that you can then use inside of Claude managed agents as part of a flow. So, I think like usually there is like quite a bit of iteration in the beginning to like figure out, okay, what are all the different phases, and how does this process actually work?

也许你需要在整个组织范围内进行事实调查，了解现有流程如今是如何构建的。然后尽可能把这些信息编码下来，再用爬山优化技能反复迭代实际的提示词并加以改进。而且这种爬山优化也可以由人来驱动。比如，你可以说：“我们来反复优化一下，看看怎么才能把这些视频做得特别好。”然后它会生成一个，你可以说——或者它可能会生成两个——然后你可以说你更喜欢哪一个。接着它会再生成两个，你再说你更喜欢哪一个，它就会这样用自己的提示词不断迭代。对。And maybe you need to do fact finding across the organization to understand how an existing process is built today. And then trying to like do your best to encode that, and then use the hill climb skill to like iterate on the the actual prompt and improve it. And that hill climb thing can be sort of human driven. So, you can say, you know, let's hill climb on making these videos it like really great. And so, it'll generate one, you can say or it'll it might generate two, and you can say which one you prefer. And then it'll generate two more, and you can say which one you prefer, and it kind of like will uh iterate on its own prompt that way. Yeah.

主持人：所以，我想问题是回到之前的话题——如果我理解错了请纠正我——回到操作系统的例子，什么时候该用 Ultra Code，什么时候该用 Claude 托管代理，比如 flows，对吧？好，好。是的，Ultra Code 适用于这种情况：你坐在笔记本电脑前，在终端里输入命令，直接以交互方式使用 Claude Code。而 Claude 托管代理的 flows 工具，则适用于你有一项需要定期重复执行的任务，并且想通过 API 将它与另一个 API 集成，以便接入你想要的任何产品界面，而不是直接使用 Claude Code。是的，我觉得你最后的描述就是正确的理解方式：Ultra Code 本质上就像是这类东西的单个实例。Host: So, I think the question was going back, tell me if I'm wrong on this, but going back to the operating system example, when would you use Ultra Code versus when would you use Claude managed agents, like flows, right? Cool, cool. Yeah, so so Ultra Code is going to be the situation where you're like sitting at your laptop, like typing into the terminal, and using Claude Code directly and interactively. The Claude managed agents flows tool would be if you have like a repeatable thing that you want to happen on some regular basis, and want to integrate an API against an API that can hook into like whatever product surface uh you want, rather than like directly just using Claude Code. Yeah, I I think that that way that you described it at the end there is is is the correct way to think about it, that Ultra Code is like the the single instance of one of these things essentially.

实际上，我们把 Ultra Code 当作一种抽象或方法论，并将其泛化成了 Claude managed agents API 上的工作流。嗯，你可以把 Ultra Code 想成是我一次性的操作，也就是我亲自和这个东西交互；而 Claude managed agents 则更像是：“好，我想把它提升一个层次，重复使用这些代理，并按我想要的触发计划来运行它们。”这是个很好的问题。问题是，如果平台上发生了恶意事件，你会如何看待责任风险？嗯，很遗憾，我不是律师。除了说我觉得这很像经营其他平台类业务之外，我没法明确对此发表太多意见。比如，你通过 API 进行集成，并在此基础上构建产品，那么归根结底，平台本身会直接与用户建立关系。And we actually took the Ultra Code like uh abstraction or methodology, and generalized it into what flows is on the Claude managed agents API. Um and so, you can think of Ultra Code as like the my one-off, like me interacting with the thing, and then Claude managed agents is like the, "Okay, I want to like level this up uh a layer, and reuse them, and um run them on whatever trigger schedule I want." And that's a great question. So, the question is, how do you think about liability if like, well, something malicious happens on platform? Um unfortunately, I am not a lawyer. Um and I will not be able to like like explicitly weigh in too much, other than um I would think of it a lot like, you know, running any other like platform-y kind of business, where if you integrate against an API, and like build a product on top of that, um at the end of the day, like the the the platform itself is going to have the relationship with with the user directly.

所以，这会有点像一种纵深防御机制：堆栈中的每一层都应该尽力确保平台上不会发生任何不好的事情。是的，问题是关于持久化事件流，以及大家在内部构建的用例。嗯，我在想有没有什么有趣的例子。有人在内部搭建了一个竞技版《卡坦岛》模拟系统，大概是利用各种事件来解释每位玩家为什么会做出相应的选择。它使用了自定义工具，这些工具是将你的 macOS 设备暴露给云端管理代理的一种方式，让代理能够在托管环境之外采取行动。嗯，我不知道，你们可能还有一些有趣的例子。And so, it's going to be kind of like a defense-in-depth kind of thing of like making like each layer of the stack should be doing their best to make sure nothing bad happens on platform. Yeah, so the question is about the durable event stream and use cases um that the folks have have built out internally. Um so, it I'm trying to think of a a fun one. Um somebody internally built out like a uh a competitive Catan uh like I guess simulation where it uses like all the events generally just to explain like why each of the players is making the choices that they're making. It uses like custom tools, which is one of the it's like the way that you would expose your Mac OS machine to uh to cloud manage agents in order to uh like take action like outside of just the hosted environment. Um, I don't know, you you might have some fun fun ones here.

嘉宾：我觉得最简单的理解方式是，假设你想自己构建一个 UI，展示所有不同的工具调用以及其他内容。我觉得很多人都在这里重复造轮子。现在有一些偏前端的工具，可以帮你暴露智能体的事件流，并以流式方式返回智能体发来的消息。不过，能够构建一个聊天界面——如今这基本算是标配了——让你可以和智能体进行对话、查看工具调用，还能点开这些工具调用。也许某个工具调用会生成一些生成式 UI，给你展示图表。尤其是在金融领域，你可能会说：“好，每次我和我的股票代码观察列表工具交谈时，只要它发现了什么有意思的东西，也把图表渲染出来，这样我就能在图表上做技术分析，或者添加一些技术指标，让它变得真正具有交互性。”这些事情用返回的事件流就能做很多。Guest: I think like the easiest way to think about it is if you want to build your own UI that has like shows all of the different tool calls and everything. So, I I think a lot of people are reinventing this the wheel here and there are some like front endy tools to help you expose your agents event streams and stream back generally like the messages that are coming back from agents, but being able to like build a chat interface, you know, it's kind of like um pretty table stakes these days, but a chat interface where you're interacting with a you know, a conversation with an agent and you can see the tool calls and then you can open those tool calls and maybe you're you're having a a tool call that has like some generative UI and it'll show you like the chart and then for especially for like the finance use case, you might say like, "Okay, every time I talk my my stock ticker watch list thing encounters something interesting, also render like the charts so I can do like my or like add technical analysis things to my chart and make it so that it's like really interactive and you could do a lot of those with like the event stream that's coming back."

嗯，不过是的，希望这回答了你的问题。我觉得是因为这里有很多事件，它是一套非常通用的工具，你可以用它来构建几乎任何 UI。Um, but yeah, hopefully that answers the question. I think there's a because there's like a lot of events, it's a very generic set of tools that you can go and use to build basically any UI on top of.

主持人：所以，问题是原生 Anthropic 和自托管之间是否存在兼容性差异。嗯，有些功能在自托管沙箱中还没有得到完全支持，不过流程是完全支持的。也就是说，在自托管沙箱中开箱即用。嗯，是的，我在想哪些功能不受支持，但总体而言，我们会尽量让各种功能在所有环境中都可用，无论是否自托管。嗯，你在创建会话时传入的环境只会改变类型。那么，问题是，在 bug hunter 这个示例中，那里有一个多仓库依赖关系图，我们实际上该如何构建——或者说，Cloud 实际上是怎么知道该做什么的？Host: So, the question is if there's any compatibility difference between like native Anthropic and self-hosted. Um, there are some features that aren't fully supported on self-hosted sandboxes, but flows are like totally supported. Like will will work out of the gate um with self-hosted sandboxes. Um, yeah, I'm trying to think what which ones aren't supported, but generally speaking we try and support like features like across all environments, whether it's self-hosted or not. Um just the environment that you pass in when you create your session will change um types. So, so the question is in the bug hunter example, where there's like a multi-repo dependency graph, how do you actually or how do we actually structure like what like how does Cloud actually know, right, what what uh what to do?

嗯，事实是，这多少取决于 Cloud。如果你让 Cloud 访问 GitHub 仓库，这些仓库最终都会挂载到 Cloud 运行所在的容器中。然后 Cloud 就可以自行决定怎么做。比如，如果你有一套很完善的 Bazel 构建配置，并且依赖树非常清晰简洁，Cloud 就可以在容器中使用任意工具来检查这些依赖。另一方面，如果你用的是纯文本语言 ABC，没有非常明确的依赖映射，或者构建依赖关系图并非易事，那么 Cloud 可能就得直接在文件中搜索符号之类的。每次进行 PR 审查时，你都需要把容器重新挂载，也就是把仓库重新挂载到容器中。对。Um and the the truth is that it's kind of up to Cloud. Um if you give Cloud access to the GitHub repositories, those will all end up mounted inside of the container where Cloud is operating. And so then Cloud can make the decision like if you have a really nice like Basil build and like have a really really crisp clean dependency tree, Cloud can use whatever tools it wants on the container in order to inspect uh those dependencies. Whereas if you have, you know, plain text language ABC that doesn't have like a very explicit uh dependency mapping or it's like non-trivial to uh build out that graph, then Cloud would probably have to like literally search files for symbols or something like that. For every PR review that you have, you would have to remount the container remount the uh repository on the container, yes.

这很像是这样：如果你想想一个普通的 CI 任务，你得先在 Git 仓库上执行某种检出操作，然后才能对它采取行动。所以问题是，如果你能提前构建出一个依赖关系图，这能帮助你优化流程吗？答案当然是能。如果你事先有某种依赖关系图的表示形式，而且你觉得生成它会花很长时间，就可以把它作为文件上传到会话中。我们支持在平台上上传各种通用文件。然后你就可以说：“嘿，Cloud，顺便说一下，你关心的完整依赖关系图在这个文件路径下。”那么，问题是，在这些多智能体拓扑结构中，嵌套层级有多少限制？It's a lot like um if how you think about like uh just a CI job generally, where that you have to like run some checkout on the Git repository um and uh then you can take action on it. So, the question is if you are able to build out a dependency graph ahead of time, will that help you optimize the flow? And the answer is totally. If you if you have some representation of the dependency graph that you know ahead of time, you think would take a long time to generate, you can upload it as a file into the session. Um like we support just general file uploads as part of the platform. And then you can be like, "Hey Cloud, like by the way, your full dependency graph you care about is at like this file path." So, the question is about how many limits of nesting are there in these multi-agent topologies?

嗯，今天真正的答案是，我们目前只支持一层嵌套。我们不支持任意复杂的巨型扇出拓扑。嗯，我们以后可能会想要支持，但我们得确保所有东西，尤其是安全方面，都真正落实到位。嗯，我先试着复述一下第一个问题，你告诉我理解得对不对。大致来说，你的机器上无法访问某种身份信息或数据，因为它位于 VPN 后面；那么，这些数据能通过任何内部 MCP 服务器访问，还是只能在你的机器上访问？好的，明白了。因为抽象这类问题的方式是使用我们称为 MCP 隧道的功能，它基本上可以让你通过 Cloudflare 隧道，为平台上的智能体提供私有访问权限，也就是私有 MCP 访问权限。Um the the real answer today is it's just one layer of nesting. We don't support like arbitrarily complex like mega fan outs of uh topologies. Um we we probably at some point will want to, but we need to make sure like everything is really uh like the safety and stuff is really locked in. Um so the first one I'm going to I'm going to try and restate it and and tell me if I'm understanding it correctly. Basically like there's some some form of like identity or data access or something that you don't have access to on your machine because it sits behind a VPN and is that data accessible via any like internal MCP servers or only on your machine. Okay, got you. Because the the way that you would abstract this would be using something we have called MCP tunnels, which is basically a way for you to give private access like private MCP access to the agents on platform using Cloudflare tunnels today.

基本上，这就像是一种让你授予访问权限、又不会把任何东西暴露到开放互联网的方式。我觉得这可能最接近你想做的事，或者说，为了让这些东西能够访问那里，你需要做的事。应该会有一些……对，对，对。是的，就是这样。连接是直接从你的私有 MCP 服务器运行的地方，连到 MCP 调用发生的地方。所以，是的，我会去了解一下这个。第二个问题只是关于 Anthropic 那边的工作方式。我们基本上非常依赖 Claude tag，让 Claude tag 自主帮我们做很多工作；比如有些工作负载会让它查看 Jira 和过去 24 小时里的所有活动，甚至主动提议修改工单之类的。And basically it's like a way for you to give access such that you don't expose anything over the open internet. And I think that's like probably the closest to what you would be trying to do or that you would need to do in order to give these things access to there would be some Yeah, yeah, yeah. It's yeah. It's It's like the connection is directly from wherever your private MCP server is running to to the place that the MCP calls happened. So yeah, I would look into that. Number the second question was on just like how how stuff works at Anthropic. We're very like Claude tag pilled basically where Claude tag does a bunch of work for us autonomously and like the these workloads where like the thing is looking at Jira and all the activity in the past 24 hours or whatever and even like proposing proactive changes to tickets and such.

我们我们有很多这类工作都是通过某种类似 Claude tag 的东西来完成的。所以，嗯，是的，是的。Claude tag 支持We we do a lot of that work through something Claude tag shaped. So um Yeah, yeah. Claude tag supports

嘉宾：是的，Claude 标签 Claude 标签可以做到这一点。Guest: Yeah, Claude tag Claude tag can do that.

主持人：我觉得你只要每天早上 6 点对 Claude 说“做这个”，我觉得它应该开箱即用。Host: I think if you just say at Claude like every every day at 6:00 a.m. like do this. I I think it would work out of the box.

嘉宾：好的。非常感谢大家的关注，谢谢。Guest: Yeah. Thank you all so much for your attention. Appreciate it.

主持人：谢谢各位。好。Host: Appreciate you guys. All right.