ClawLabs：从个人编码走向共享智能体

简要说明Brief Description

ClawLabs 团队介绍了他们如何从个人编码会话转向持久化协作代理，演示了共享产物和社区工作流，并推出部署控制功能和一个开源代理平台。The ClawLabs team explains its move from individual coding sessions to persistent collaborative agents, demonstrates shared artifacts and community workflows, and introduces deployment controls and an open-source agent platform.

目录Table of Contents

• 从终端会话到协作式智能体
  From Terminal Sessions to Collaborative Agents
• 一起制作游戏
  Building a Game Together
• 审查意图与协作工作流
  Reviewing Intent and Collaborative Workflows
• 产物、提示词与代理驱动的设计
  Artifacts, Prompts, and Agent-Driven Design
• 持续性维护与共享实验
  Persistent Maintenance and Shared Experiments
• 部署控制与扩展代理
  Deployment Controls and Scaling Agents
• 权限与社区智能体演示
  Permissions and Community Agent Demonstrations
• ClawLabs 与开源基础设施
  ClawLabs and Open-Source Infrastructure

从终端会话到协作式智能体From Terminal Sessions to Collaborative Agents

我们来聊聊吧，今天的分享会非常精彩。我会向你展示这一年来维护开源项目时学到的东西、我的工作流程是如何演进的，以及我们把 OpenClaw 的开发迁移到 OC 本身后发生了什么。现在，我们稍微聊聊用来和代理交互的工具框架。Let's talk, we have a really exciting session today. I'm going to show you what I learned over the year maintaining my open source projects, how my workflows evolved, and what happened when we moved development of OpenClaw to OC itself. Now, let's talk a little bit about the harnesses we use to talk to our agent.

整整这一年，这场变革都是从 CLI 开始的。去年，我完全生活在终端里。分屏，开着一大堆 Codex 会话，我的代码在左右两边飞来飞去。This whole year, this whole revolution started in the CLI. Last year, I was fully living in the terminal. Split screen, lots of Codex sessions, my code was flying around left and right.

随后，随着时间推移，模型变得越来越好、越来越聪明。而我需要不断按 Esc 来引导智能体的情况也越来越少。现在基本上已经不需要这么做了。And then over time, the models got better and smarter. And the time where I had to smash escape to steer the agent decreased. It's not really needed anymore.

很明显，我的工作流程需要改变。之后我转而使用 Codex Desktop。现在聊天应该在那边。It was clear my workflow needed a change. Then I moved to Codex Desktop. Now the chat should be there.

这使得同时启动大量会话变得容易得多。It made it far easier to start a large number of sessions concurrently.

这段文本更易于阅读。它具有可视化效果，还内置了浏览器。The text is easier to read. It has visualization. It has an inline browser.

它真的有代码审查。我都不知道。真不错。It has really code review. I don't even know. It's nice.

现在运行所有这些会话和实验变得如此轻松，我的电脑却快撑不住了。磁盘空间用完了，风扇响得像喷气式发动机。Now that it's so easy to run all these sessions and experiments, my machine was dying. I ran out of disk space. My fans sounded like a jet engine.

于是我开始使用多台电脑来处理更多工作。我把 Codex 连接到了 DevBoxes，这样就能分散负载。So I started using multiple computers to run more work. I connected Codex to DevBoxes. So I could spread the load.

我搭建了 Grabboxes，这样代理就能在云端运行测试。我想我会查看——不知道，六台或十台机器，它们正忙着处理我所有的代码和实验。有时是在一个 Codex 里，有时则用 VNC。And I built Grabboxes so agents would run the tests in the cloud. I think I'd peek ahead, I don't know, six or ten machines that were crunching through all my code and my experiments. Sometimes in one Codex and sometimes they used VNC.

后来我意识到，这是所有这些机器都无法解决的问题。当队友们想要参与进来时，他们还是得通过我。我们都在使用智能体，但实际上并没有真正协同工作。And then I realized that was a problem all these machines wouldn't solve. When my teammates wanted to get involved, they still had to come through me. we were all working with agents, but we weren't really working together.

顺便说一句，我很喜欢这张幻灯片。这就是 token 的样子吗？Frank Allenby 今年创造了“肉身代理”（meat proxy）这个说法。我说的就是这个。By the way, I love this slide. It's that's how tokens look? Frank Allenby, he coined this term this year, meat proxy. what I'm talking about.

咱们别做个肉身代理。也就是你充当 AI 的接口，在双方之间来回传递消息。那团队里的其他人怎么参与进来呢？Let's not be a meat proxy. That's when you are the interface to AI and you pass messages back and forth. So how does the rest of the team get involved?

你如何看清它在做什么、理解它的决策，并帮助推动工作进展？这正是我们开始用 OC 构建 OC 时着手探索的问题。对于还没用过它的各位，OpenClaw 是一个用于构建持久运行智能体的开源平台。How do you see what it's doing, understand the decisions, and help move the work forward? That's what we started exploring as we began using OC to build OC. And for those of you who haven't used it, OpenClaw is an open source platform for persistent agents.

它可以在你的各种工具、设备和团队中使用。你可以使用订阅账号登录。现在还有一个很酷的“使用 ChatGPT 登录”按钮。It works across your tools, your devices, your teams. You can sign in with your subscription. There's now a fancy sign in with ChatGPT button.

它会识别你的插件，你甚至还能看到自己的 Codex 会话。另外，我们来谈谈治理。It picks up your plugins, and you can even see your Codex sessions. Also let's talk about governance.

OpenClaw 是开源项目，由独立的非营利组织 OpenClaw Foundation 所有。OpenAI 慷慨地支持我们的使命。我在 OpenAI 的团队致力于探索如何设计实用的智能体、如何大规模运行它们，以及它们如何帮助我们改善工作和个人生活。OpenClaw is open source, and it's owned by the OpenClaw Foundation, an independent nonprofit. And OpenAI generously supports our mission. My team at OpenAI explores how to design agents that are useful, how to run them at scale, how they can help us improve both our work and our personal lives.

最终，这也关乎自动化软件工程师可能会是什么样子。我的论点是，模型进步得太快、能力越来越强。我们的工作方式、工具链的运作方式，都跟不上它们的步伐。And ultimately also how an automated software engineer could look like. My thesis is that the models are getting so good so fast. The way we work, the way our harnesses work, we just can't keep up.

一年前甚至半年前还合理的工作流，如今已经让人觉得受限而过时。现在，我的工作之一就是探索未来的发展方向。要不要一起看看？Workflows that made sense a year ago, or even six months ago, already feel constrained and outdated. Now, part of my work is to explore where things go. Should we take a look?

你们已经看到，今天演示之神眷顾着我们。那我们来看看 PartyClaw 是否在线。嘿，PartyClaw，你还活着吗？You already saw that the demo gods love us today. So let's see if PartyClaw is online. Hey PartyClaw, are you alive?

嗯，语音翻译能用，尽管不太完美。Well, the voice translation works, even though it's not perfect.

一起打造游戏Building a Game Together

太好了。那么，我要做一个游戏。那个青蛙要穿过车来车往的马路的游戏叫什么来着？你能帮我做出来吗？目前为止没什么新意，对吧？Perfect. So, I'm going to build a game. what's this game called where the frog has to go over a street with cars? Can you build it for me? so far it's nothing new, right?

到目前为止，你的智能体会构建这个项目，最终会给你返回一个拉取请求。然后，你可能会在 Codex 中打开这个拉取请求并继续处理。我认为有更好的办法。So far your agent would build this, and eventually it will get you a pull request back. And then you probably would take that pull request and open it in your Codex and keep working on it. I think there's a better way.

哦，对，Frogger。哎呀，我真傻。它已经找到一个我们可以用的版本了。Oh, yeah, Frogger. Oh, I'm so silly. It already found a version we can use.

莉莉·莱恩。好的。是谁创作了莉莉·莱恩？Lily Lane. Okay. Who made Lily Lane?

我们点一下那个。好吧。该死。Let's click on that. Okay. Well, damn.

我正在使用 ultrafast。它快得惊人。但那比 ultrafast 还快。I'm using ultrafast. It is amazing. But that was faster than ultrafast.

谁做的？好吧，凯文和我的团队已经把游戏做好了。哦，那就看看我能不能在台上拿自己开个玩笑。Who did that? Okay, Kevin and my team already built the game. Oh, well, let's see if I can make a joke out of myself on stage.

我能应付得来吗？Will I manage?

哎哟。喔！喔！Oof. Whoop. Whoop.

哦，天哪。每次看演示我总会漏掉那个。不过……啊。Oh, God. in all the demos, I always miss that. But... Ah.

我其实不太想让它变成青蛙。能不能变成龙虾？我也不知道。I don't really want this to be a frog. Can this be a lobster? I don't know.

不！看吧，我更快。你看那些正在打字的人。No! See, I was faster. You see all those people typing.

但第一个提示是我做的。哈哈，不应该是大象。But I did the first prompt. Ha, ha. Should not be an elephant.

我们来做一道龙虾吧。啊，看看这要花多长时间。We'll do a lobster. Ah. Let's see how long that takes.

好吧，趁我们等着的时候，主题演讲里有好多很酷的内容。嘿，我现在就在 OpenAI DevDay 现场参加一个工作坊，我想我错过了一小段主题演讲。OpenAI 在 2026 年 DevDay 上都发布了哪些很酷的东西？Well, while we wait, there are so many cool things that were in the keynote. Hey, I'm just at OpenAI DevDay here and doing a workshop, and I think I missed a little bit of the keynote. What are all the cool things that OpenAI released in DevDay 2026?

你能帮我做一个仪表盘吗？好的。趁它运行的时候，我们要不要回头看看《青蛙过河》？Can you make me a dashboard? All right. While it works, should we check back on Frogger?

好吧。我还看到有三个人在线，人数不多。All right. And I also see there's three people online. That's not a lot of people.

我们要不要把气氛搞得更像派对一点？Should we make this a bit more of a party?

好耶！行。我想我得点击这个按钮。Yeah! Okay. I think I have to click the button.

所以，如果你动作很快，可以使用这个 URL，然后加入我，一起制作一些有趣的小游戏。它在这里是做什么的？这是专为所有人都拥有相同访问权限的团队设计的。So if you're very fast, you can use this URL and then join me to make silly games. What is it doing here? So this is designed for teams that all have the same access.

想象一下，你在做开源项目，并且想和团队协作。过去我们通过拉取请求来协作，但现在感觉这已经不是未来的发展方向了。所以在这里的服务器上，任何人都能看到彼此的会话。Think about you doing open source and you want to work with a team. And we used to have pull requests to collaborate, but that just feels not the future anymore. So here on the server, anyone can see each other session.

你可以向彼此的会话发送提示。这项功能是为协同工作的人设计的。不要邀请你不希望其干扰你会话的人。You can prompt each other session. It's meant for people that work together. Don't invite anyone who you don't want to mess with your sessions.

未来我们在构建 OpenClaw 时，会用这个来邀请一些顶尖贡献者。你们获得代币，我们也能汲取一些你们的想法。我觉得这是笔非常划算的交易。整个工作坊和这个想法，都是为了启发你们思考。We're going to use this in the future as we build OpenClaw to invite some of our top contributors. you get the tokens and we get a little bit of your ideas. I think it's an amazing deal. And this whole workshop, this idea, this is really to get you thinking.

在这个智能体时代，还有哪些工作流程我们现在仍在用老办法处理，其实可以做得好得多？哦，我觉得我的龙虾游戏准备好了。哦，现在变成龙虾和鲨鱼了？Which other workflows do we currently do in the old way that we could do so much better in this agentic age? Oh, I think my lobster game is ready. Oh, it's lobsters and sharks now?

好吧。哦，它死机了。该死。Okay. Oh, it died. Damn it.

好吧，你们都在这儿玩得开心的时候，天哪，我们到底放进来了多少人？伙计们，咱们得关上大门了。是啊，你们已经看到那些事正在发生了。Well, while you all are having fun here, and oh, my God, how many people did we let in? We need to close the gates, guys. Yeah, you already see that stuff's happening.

我们之后会再来检查这件事。现在切回幻灯片。刚才你们看到我们一起稍微做了一会儿游戏。We're going to check back later on this. Let's switch back to the slides. So, you just saw us working a little bit together on a game.

几个月前，我们开始尝试智能体参与多人协作会是什么样子。这改变了我们的团队协作方式。不仅提高了工作效率，也让我们更开心了。如果能看到彼此的提示词，就能避免很多冲突。Now, we started experimenting with how would multiplayer work with agents a few months ago. And it has changed how we work as a team. Not only has it increased velocity, it also just made us happier. there's so many conflicts that you can avoid if you see each other's prompts.

而现在感觉我们确实更像一个团队在协作了。And now it feels like we're actually more working like a team.

意图审查与协作工作流Reviewing Intent and Collaborative Workflows

今天早些时候，你们还看到了 Dots——这是 OpenAI 面向企业与智能体协作体验的一项非常酷的官方集成。OpenClaw 探索了一些相同的理念：让智能体与整个团队的人并肩协作。And earlier today, you saw dots, which is a really cool first-party integration of the OpenAI experience for companies to work with agents. OpenClaw explores some of the same ideas. Agents working alongside a whole team of people.

OpenClaw 让你可以自由选择模型、修改代码，还能完全在自己的基础设施上运行。我认为我们应该为此喝彩。我很高兴 OpenAI 支持所有这些不同的做法。OpenClaw gives you the freedom to choose your model, change the code, and you can fully run it on your own infrastructure. And I think we should celebrate that. I love that OpenAI is supporting all these different approaches.

而且，开源社区可以将这些想法拓展到我们甚至未曾设想过的方向。当然，我们也在认真思考如何让你的 Dots 和 Claw 互相通信。还有一件我刚才没提到，那就是你的提示词。And that the open source community can take these ideas in directions we haven't even imagined. And of course, we're really thinking about ways how your dots can talk to your claw and vice versa. Now, one thing I didn't talk about is your prompts.

现在，一句话就能生成数千甚至数万行代码。但背后的对话往往仍留在某个人的本地机器上。A single sentence can now produce thousands or tens of thousands of lines of code. But the conversation behind it often stays on someone's local machine.

我审查过很多 PR，知道改了什么，却不知道为什么。给我看看提示词。于是我试了这个做法：添加了一项技能，让大家发送一份脱敏后的对话记录，并将它附在 PR 中。I reviewed so many PRs where I understood what changed, but not why. show me the prompt. So I tried this. I added a skill that asked people to send a sanitized version of the transcript and attach it to the PR.

很少有人答应。分享自己如何与智能体交流，感觉出乎意料地私密。说真的，就连我有时也会拒绝。And very few said yes. Sharing how you talk to your agent feels surprisingly personal. Heck, even I sometimes said no.

我不知道。你知道吗？所以，当我们让整个团队都能看到这些会话时，确实有一天感觉挺别扭的。I don't know. You know? So when we made sessions visible to the whole team, it did feel awkward for a day.

我们读着彼此的对话，轻轻笑了几声。然后，这就成了新的常态。We were reading each other's conversations. We laughed a little. And then it became the new normal.

现在我们可以相互学习，也能更早发现错误。让我给你看看它是什么样子。And now we learn from each other. And we catch mistakes earlier. Let me show you how it looks like.

所以，演示之神依然站在我们这边。我想象自己醒来时，还没喝咖啡。So, the demo gods are still on our side. And I imagine I wake up. I didn't have my coffee yet.

我看到了 Roboclaw 提交的这个拉取请求，然后浏览了一下。And I see this pull request from Roboclaw. And I scroll through it.

然后，有人改了表格的样式，把它们缩小了。可我做的也是同样的工作，而且就在前一天，我还把它们放大了。感觉我的皮质醇水平都要升高了。And there, somebody changed how tables look and made them smaller. But I did the same work. And I made them bigger just the day before. like my cortisol level would go up.

这是谁干的？我是说，为什么？维克多做了什么？It's who did this? Like, why? What did Victor do?

或许过去我会给他发一条不太加州风格的消息。但现在，我可以点开工作会话，看看他做了什么，对吧？And maybe in the past, I would have dropped him a not very Californian message. But now, I can click on the work session. And I can just check out what he did, right?

而且看起来他确实复用了我的会话。或者他甚至还给我发了消息。我应该多看看邮件。And it actually looks like he reused my session. Or he even pinged me. I should read my emails more often.

我喜欢这个。看到了吗？我们刚才还在聊我们是怎么和代理沟通的。I love this. See this? We were just talking about how we talked to our agents.

而且维克多显然对我们的 CI 耗时很不满意。光讲 CI，我大概就能讲满一整场。还有一点是，我们目前的工作方式还没有跟上这个新时代所需要的一切。And Victor clearly is not happy about our CI times. I could probably fill a whole talk about just CI. And that it's another point of the way we currently work did not catch up with all the things that we now need in this new day and age.

制品、提示词与智能体驱动的设计Artifacts, Prompts, and Agent-Driven Design

我认为 CI 按照目前的情况已经完全说不通了。我现在正在试用 Codex。也许如果加入一些智能能力，就能更容易地判断哪些测试子集真正有用。And I think CI, as it currently doesn't make any sense anymore. I'm currently experimenting with Codex. Maybe if we add intelligence, it could be more easier to decide which subset of the tests are actually useful.

不过我觉得这本身就值得单独聊聊，也许可以留到改天再说。但如果我继续往上翻，实际上看起来维克多为此花了不少心思。我错了。But I feel that's a talk in itself and maybe something for another day. But if I keep scrolling up, it actually looks like Victor put a lot of thought into it. I was wrong.

这其实真的很棒。他基本上就是拿了我的成果，然后把我的东西修好了。太赞了。This is actually really good. He basically took my work and then just fixed my stuff. This is awesome.

你知道吗？午餐危机总算化解了。现在，我带你简单参观一下。You know? Lunchtime crisis averted. Now, I give you a little bit of a tour.

我们觉得，服务器与其说是一个应用，不如说更像一个地方。构建项目时会涉及更多事情，不只是代码。因此，侧边栏里自然会有很多应用。We see our server less like an app but more like a place. there's a lot more that goes on when you build a project. It's not just about the code. So naturally, we have a lot of apps in the sidebar.

有些是仪表板，有些是插件。And some are dashboards and some are plugins.

当然，也没什么新东西。就像我们以前有的那些一样。它们只是散落在不同的服务器上。Of course, there's nothing new as well. Like we used to have them. They were just scattered around on different servers.

有些人使用 GitHub Actions 进行更新。那里有数据库，也有不同的代码仓库。Some used GitHub actions to update. There were databases. There were different repositories.

就像我都不知道我们大家用的是哪一个。显然，让团队来更新这些内容并不是很有吸引力。比如，我给你举个简单的例子。Like I didn't even know which one we all had. And clearly, it was not very inviting for the team to work on an update on these. Like I give you a simple one.

既然我们刚才在聊 CI，我就做了一个仪表盘，用来跟踪 CI 耗时情况。显然，我们还需要做更多工作。Because we were just talking about CI. So I made this dashboard that tracks how our CI times are. And clearly, we need to do more work.

但妙处在于，每当我想调整仪表盘时，只要告诉代理：“把这个改成蓝色。”它也不会因此对我有意见。然后我们就可以边做边更新仪表盘。But the beauty is, any time I want to tweak the dashboard, I can just tell the agent, make this blue. And it will not like me about this. And then we'll just update the dashboard as we go along.

那么，大多数仪表板究竟是什么？它就像一个会话，能够生成某种产物。然后再由一个自动化流程来更新这个会话。Now, what are most dashboards really? It's like a session that generates some artifact. And then there's an automation that updates that session.

当然，这个制品里包含代码，所以从某种意义上说，它是软件。不过，仔细想想，现在有个更合适的词。And of course, the artifact contains code. So, it's software in a way. But there's a better word now if you think about it.

我们在这些事情上的迭代速度太快了。它还是软件吗，还是说该叫果冻件？我喜欢果冻件。We iterate so fast on things. Is it still software or is it jellyware? I like jellyware.

接下来我再给你看几样东西。当然，我们还有社区。来吧。So let me show you a few other things. Of course, we have the community. Here we go.

我们会跟踪所有这些数字。我们有多少个 PR？报告了多少个问题？We track all these numbers. How many PRs do we have? How many issues are reported?

合并了多少个？有些人真的很喜欢图表。我们在服务器上也会做些傻事。How many are merged? Some people are really into graphs. We also do silly things on the server.

就像每日 Claw 一样，它试图理清服务器上汇总的所有内容，然后把它们整理成一份报纸。有时候做得还真挺不错的。Like the daily Claw that tries to make sense of all the things that merge on the server. And then makes a newspaper out of it. And sometimes it's actually quite good.

但很多时候它都大错特错。不过，关键是你在创作时得享受其中的乐趣。如果想做好工作，这才是全部意义所在。But many times it's very wrong. But it's just, you have to have fun when you build. That's the whole point if you want to do good work.

我也有一个低质内容检测器。I also have a slop meter.

因为如果你们都和智能体一起工作，很容易就会让智能体生成越来越多、越来越多的代码，直到把事情搞得一团糟。所以你应该始终尽量推动智能体清理代码。正如你们所看到的，自从我开始做这个仪表板以来，我确实提高了大家的意识。Because if you all work with agents, it's very easy to like agents generate more and more and more and more and more code until you're like in a bad place. So you should always like try to push the agent to clean up. And as you see, ever since I started the dashboard, I definitely raised awareness.

而且进展正朝着正确的方向发展。这些都只是 HTML 插件。我们还有一些是真正的代码。And things are moving in the right direction. Now these were all just HTML plugins. We also have some that are actual code.

比如你去看报表的话。我是个好奇心很强的人，所以我总想看看大家都忙成什么样。Like if you go on reports. I'm a curious person. So I always want to see how busy is everyone.

正如你所看到的，这里展示了他们的活动情况：提交了多少次代码、在 Discord 上有多活跃。我们还构建了会议功能。看起来现在正在进行一场会议。And as you can see, the activity, how many commits do they do, how active they are on Discord. Now we also built meetings. And it looks like there's a meeting in progress right now.

我们以前有自己的软件，也一直为它付费，后来我就想：我们为什么要为这个付费？We used to have our own software that we paid for until I was why do we pay for this?

这是一个提示词。所以我们刚刚把它变成了提示词。它使用 Luna 来总结我们的转录稿。This is a prompt. So we just made it a prompt. It uses Luna to summarize our transcripts.

而且每次我们开会时，这个智能体都会直接加入我们的 Discord。最酷的是——不过今天我就不演示了，因为语音功能已经把演示之神惹恼了——我们的智能体现在还能和我们对话。不只是记笔记，那早就不新鲜了。And the agent just joins us in Discord whenever we have a meeting. And the coolest thing, but I'm not going to show you that today because we already angered the demo gods with voice, is our agent can now talk back. like just taking notes. That's nothing new.

但连接 GPT Live。然后，你就可以向智能体询问与你的工作相关的问题，或者另行启动新工作。因为代表服务器的是同一个身份，而且主会话也掌握所有信息。But connect GPT Live. And then you can actually ask the agent about your work or spin off new work. And because it's the same identity that represents the server and the same main session that knows everything.

效果非常好。现在，我们来看看实际上有哪些人在线。如果我稍微往下看一点。This works really well. Now, let's see who is actually online. If I go down there a little bit.

我很喜欢这个视图，也总是好奇 Victor 在忙些什么，因为他通常会把我的代码改得更漂亮。I love this view. And I'm always curious what Victor is up to. Because he usually takes my code and makes it pretty.

看来 Victor 正在做一些表情符号相关的东西。So it looks like Victor is working on some emoji stuff.

但我知道他今天早上给了我一次机会。他告诉我，他把侧边栏做得更好看了。比如，我加了一个小视图，这样你就能看到谁在线、谁正在运行。But I know he gave me a shot this morning. He told me that he made the sidebar nicer. Like I added this little view so you could see who is online, who is running.

但它看起来不太好。所以他跟我说他做得更好，然后让我往下滑着看。But it doesn't look very good. So he told me he does it better. And then let me scroll through.

很棒的一点是，你真的能看到他是怎么工作的，也能看到他如何与自己的智能体打交道。有时候，他用词还挺丰富多彩的。And the cool part is you can actually see how he works. And just seeing how he approaches his agents. And sometimes he uses very colourful language.

但他居然能做出这么多不同的设计迭代。我再往下滚一点，让你看看。看这个。But he manages to squeeze out so many different design iterations. Let me scroll a little bit more so you see. Look at this.

他为此专门开发了整套编辑器来进行调整。可他还是不满意。我觉得维克多真的很难取悦。He built whole editors to tweak this. And then he was still not happy. I think it's very hard to please Victor.

哦，然后又有人加入进来。接着他们一起合作完成这件事。还有更多版本。Oh, and then somebody else was joining in. And then they were collaborating together on this. There's more versions.

我甚至不知道我们更新的是哪一个。I don't even know which one we updated.

但最终，他们找到了一个看起来非常有说服力的东西。顺便说一句，所有这些可视化效果在 Codex 应用中也能使用。所以如果你还困在 CLI 里，是时候升级了。But eventually they found something that looks very compelling. And by the way, all these visualizations, they work in the Codex app as well. So if you're still stuck in the CLI, it's time to upgrade.

你还困在2025年呢。更妙的是，我知道有些公司会开着终端参加视频通话，然后用它来跟自己的智能体交流。You're stuck in 2025. Even better, I know some companies that jump on a video call with a terminal. And then use that to talk to your agent.

还有一种更好的方法。现在，我想先向你介绍一下这里更宏大的愿景。归根结底，我们要打造的是一个能够构建我们软件的智能体。There's a better way. Now, I want to give you a little bit of the broader vision here. Ultimately, what we are building here is an agent that builds our software.

这有点像通往软件工厂的路径。有些工作必须亲自动手。你希望直接与智能体结对协作。This is kind of like the way to the software factory. And some of the work has to be very hands-on. You want to pair directly with an agent.

但我也会运行很多循环。因此，我用它们来监控反馈，然后自动提交修复。有些循环还会在不同的机器上运行。But I also run a lot of loops. So I use them to monitor feedback and then automatically land fixes. And some of those loops run on different machines.

持续维护与共享实验Persistent Maintenance and Shared Experiments

有时它们会停止运行，或者需要调整。有时我甚至不确定它们在我的所有机器上的什么位置。这些循环需要维护。And sometimes they stop or they need adjustment. And sometimes I wasn't even sure where on all my machines they were. Those loops need maintenance.

还有，归根结底，我们还能在哪些方面创造价值？关键在于品味。品味尤其难以描述。And also, ultimately, where do we still bring value? It's about taste. Taste is uniquely hard to describe.

这往往只是一种感觉，而我不想把这件事完全交给我们的智能体。因此，这在团队中并没有真正奏效。It's often just a feeling. And that's not something I want to fully entrust on our agent. So this didn't really work well in the team.

比如，我当时运行的循环太多了。于是我们正慢慢地、一个接一个地把这些循环迁移到团队服务器上，让其他人也能学习。因为归根结底，维护那个构建我们软件的智能体的是我们。Like, I was running too many. And slowly, one by one, we're moving all these loops to our team server so other people can learn. Because ultimately, we maintain the agent that builds our software.

那么，我们运行哪些类型的循环？它们是如何工作的？比如，我有一个监控程序会扫描 Discord、Twitter 和 GitHub。我构建了一些小型 CLI 工具。Now, what types of loops do we run and how does that work? Like, I have a monitor that scans Discord, Twitter, and GitHub. I built little CLIs.

因此，智能体可以很轻松地获取大量数据。然后我们对这些数据进行分类。So this is very easy for the agent to get a lot of data. And then we categorize them.

对于某些类别的问题，比如更新问题，智能体会阅读报告并进行调查。它会启动一台新的测试机器来复现错误。如果能够明确复现，就会提出修复方案。And for certain categories, like update issues, the agent will read the report, will investigate. It will spin up a new test box to reproduce the error. If it can clearly reproduce it will come up with a fix.

它会检查问题是否已修复。如果这些都能正常运行，还会有另一个代理来审核这项工作。审核通过后，你的工作就可以自动合并。It will see if it's fixed. And if all of that works, there's going to be another agent that reviews that work. And then your work can automatically land.

那么，循环和目标有什么区别？循环基本上会持续检查。这是我们不断进行的一件事。Now, what's the difference between a loop and a goal? So the loop basically keeps checking. This is something we do continuously.

关于性能，我还有一个问题。我们每天都会发布一个新的团队服务器版本。然后我的更新循环会检测到这一点。I have another one for performance. We do another release of the team server every day. And then my update loop will detect that.

我们会进行快照，然后看看是否有任何退化之处，或者哪些地方需要改进。现在，如果我执行 /goal，这项工作就会朝着一个明确的最终状态推进。比如，我目前的目标是逐步将代码库从同步 SQLite 迁移到异步数据库访问。We'll take snapshots and we'll see if there's anything that regressed or it needs improvement. Now, if I do, slash goal, that work is towards a defined end state. Like right now, I have a goal that slowly, bit by bit, migrates the code base from synchronous SQLite to async database access.

我们这里仍在使用 SQLite，但随着服务器规模扩大、我们做的事情越来越多，我们正在考虑让数据库本身也可以替换。它已经运行了18天，目前仍在我的本地 Codex 上运行，到现在已经落地了100多个 PR。We're still using SQLite here, but as the server grows and as we do more, we're thinking about maybe making the database itself swappable. So this has been running for 18 days. It's still running on my local Codex and it has landed over 100 PRs so far.

随着时间推移，我希望每个项目都能配备一个智能体。它应该拥有自己的记忆、自动化流程、目标和循环机制，而我可以向它下达高层指令，让它负责维护和改进我的软件。那么，我们现在要不要回顾一下大家目前都在构建什么？Over time, I want an agent per project. It should have its own memory, automations, goals, loops, and I can give it those high-level commands and it maintains and improves my software. Now, should we check back what people are building so far?

谁有幸真正用过在线服务？我看到有几个人举手了。对，你们反应真快。Who got lucky enough to actually use online? I see a few hands. Yeah, you were fast.

那么开始吧。好，我们这里有新歌雷达。我们有……So here we go. And okay, we have the release radar here. We have...

哦，这是个很酷的游戏。好吧，你不介意我吧。Oh, this is a cool game. Okay. You don't mind me.

我正要……你是什么意思？应该让代理来玩吗？是的。I'm just going to... What do you mean? Should the agent play? yes.

到目前为止，智能体可能已经更出色了，尤其是因为这些都支持 CUA。By now, agents are probably better, especially because this all supports CUA.

哦，岛屿建造游戏。好吧，这是什么？Ooh, island builder. Okay. What is this?

我完全不知道这是怎么运作的。接下来我们该确保什么？好吧。I have no idea how that works. What should we assure next? Okay.

和 Claw 一起玩你画我猜。给这个概念起个名字。还有人能帮帮我吗？Play Pictionary with the Claw. Name the concept. Can anyone else help me?

谁在线？帮帮我。给我个提示。Who's online? Help me. Give me a hint.

不管怎样，我会继续让你在 Explorer 里体验。如果你没拿到名额，也可以下载它并部署到自己的服务器上。再说一次，它完全兼容 Codex。Well, anyhow, I'll keep letting you play in the Explorer. And if you didn't get a spot, you can just download it and set it up on your own server. Again, it fully works with Codex.

我们确保只需点击一次，就能将订阅一并纳入，然后你就可以一起参与了。我们一直在向大家展示小团队如何与智能体协作。随着越来越多的团队开始使用智能体，我们又遇到了一个问题。We made sure it's a one-click to get your subscription included and you can play along. So, we've been showing you how a small team can work with agents together. And as more and more teams started using agents, we ran into another question.

如何在整个组织范围内运行、配置并保障它们的安全？我们对此深有体会。这就是我们构建企业级控制平面的原因。How do you run, configure, and secure them all across an organization? And we definitely felt that pain. That's why we built an enterprise control plane.

而今天，我们将其开源。And today, we are open sourcing it.

部署控制与智能体扩缩容Deployment Controls and Scaling Agents

接下来，我请 Kevin Lin 上台。他将向大家展示我们打造的成果。嗨，Kevin。And with that, I'm going to bring up Kevin Lin. He's going to show you what we built. Hey, Kevin.

Hi, everyone. I'm Kevin Lin. And I lead OpenClaw Enterprise at OpenAI.

现在，OpenClaw Enterprise（简称 OCE）为 OpenClaw 引入了一个开源控制平面。它让 OpenClaw 支持多租户，并为智能体部署增添治理能力和严格的安全边界。我们认为，安全的运行时环境是让每个人都能使用 dots 和 OpenClaw 这类持久化智能体的必要基础。Now, OpenClaw Enterprise, or OCE in short, introduces an open source control plane to OpenClaw. It makes OpenClaw multi-tenant and adds governance and hard security boundaries to agent deployments. Now, we believe that a secure runtime is a necessary foundation for bringing persistent agents like dots and OpenClaw to everyone.

因此，今天我将带大家了解我们内置于平台中的一些安全基础机制。首先，我们在受信任和不受信任的工作负载之间强制实施严格的隔离。这意味着确定性代码在独立的容器中运行；在我们的场景中，它还与智能体运行所在的集群分离。And so, today, I'll walk you through some of the security fundamentals that we've built into the platform. First, we enforce hard boundaries between trusted and untrusted workloads. This means that deterministic code runs in a separate container, and in our case, a separate cluster from where the agent is executing.

这可以防止代理更改其运行环境。其次，我们通过沙箱机制限制代理的操作，包括控制其网络访问权限、可写入的文件，以及可以推送代码的代码仓库。This prevents agents from changing their runtime. Second, we constrain what agents can do with sandboxing. We control their network access, what files they can write to, and what repositories they push to.

这样，我们就能为智能体提供实用工具，而不必让它们访问所有内容。接下来，我们在现有控制措施之上加入基于模型的审查。我们使用自动审查功能，它在 Codex 和 OpenClaw 中均可用，通过另一个模型审查智能体的操作，并在可能危险的操作执行前将其拦截。This lets us give agents useful tools without giving them access to everything. Next, we add model-based reviews on top of our existing controls. We use auto-review, which is a feature in both Codex and OpenClaw that uses another model to review the agent's actions and block potentially dangerous ones before they execute.

最后，智能体以各自的身份运行，拥有专用服务账户和可配置的连接器。这意味着智能体不会冒充用户，而是独立且持久存在的实体。And finally, agents run as their own identities with dedicated service accounts and configurable connectors. This means that agents aren't impersonating users. They are themselves persistent entities.

它们可以被独立控制、审计，并在必要时关闭。They can be independently controlled, audited, and if needed, turned off.

权限与社区代理演示Permissions and Community Agent Demonstrations

正是这些控制措施，帮助我们将内部 OpenClaw 部署规模扩大到 400 多个 Claw。它们全天候运行，修复代码、进行实验、分析数据，并上报安全问题。接下来，我来给大家展示一下这是什么样的。Now, it's these controls that have helped us scale our internal OpenClaw deployments to over 400 Claws. They work around the clock, fixing code, doing experiments, analyzing data, and escalating security concerns. So now, I'm going to show you guys what that looks like.

这是个很难的你画我猜游戏。我不知道那是什么。好吧。This is a hard Pictionary game. I don't know what that is. Okay.

由于我们今天要向世界推出 OC，让我们创建一个社区代理。你现在看到的是预设。预设是可配置的代理套件，组织可以对其进行配置，从而让代理按照你们的安全最佳实践启动。Because we are unveiling OC to the world today, let's create a community agent. What you're seeing right now are presets. Presets are configurable agent bundles that an organization can configure so that agents are launched with your security best practices.

我们把这个智能体叫作 Fred。我们要用 Codex harness 为 Fred 配置环境，并给他一个服务账户。Let's call our agent Fred. For Fred, we're setting up with the Codex harness. We're going to give him a service account.

我们将向 Fred 提供代码库的只读访问权限，以便他能够回答问题。We're going to give Fred read-only access to the code base so that he can answer questions.

我们的路线图放在 Linear 上，所以我们给 Fred 开通 Linear 的访问权限。不过要确保，任何写入 Linear 的操作都必须经过团队审批。We keep our roadmap on Linear. So let's give Fred access to Linear. But let's make sure that if we're writing to Linear, that those go through team approval.

接下来，我们确认一下 Fred 是否在 Slack 中。我已经有一个 Slack 应用和机器人令牌。由于这是一个预设，Fred 已经配置好了所需的频道。Next, let's make sure that Fred is in Slack. I already have a Slack app and bot token. And because this is a preset, Fred is already configured with the necessary channels.

那么，我们来让 Fred 出场。现在 Fred 正在启动，我想切换到我们另一个智能体 Cooper。Cooper 是一个软件智能体。So let's bring Fred forth. Now, while Fred is coming online, I want to switch to another agent that we have, Cooper. So Cooper is a software agent.

把它看作一位长期协作的队友。他帮助我们梳理问题优先级、修复 bug，并端到端地合并功能。现在，OCE 控制台虽然能用，但看起来有点朴素。And think of him as a persistent teammate. He helps us triage issues, solve bugs, and merge features end to end. Now, the OCE console, it's functional, but it's a bit plain.

那我们来点新鲜的。Cooper，你能在仪表板上添加一张按代理划分的 Token 使用量热力图吗？如今，软件代理是我们部署的最受欢迎的代理类型之一。So let's spice it up. Cooper, can you add a heat map of token usage per agent to the dashboard? Now, software agents are one of the most popular agents that we deploy.

在 OpenAI 内部。事实上，如果你在过去一个月内提交过 Codex 问题，很可能有一只 OpenClaw 对其进行了分诊、复现，并且在很多情况下端到端地完成了合并。好吧。inside of OpenAI. In fact, if you've submitted a Codex issue any time within the last month, there's a good chance that an OpenClaw has triaged it, reproduced it, and in many cases, merged it end to end. All right.

所以 Cooper 正在发挥作用。Peter 今天谈到的很多内容都与未来的工作方式和多人协作有关。每个通过 OCE 部署的智能体都内置了这些功能。So Cooper is working. And a lot of what Peter talked about today is the future of work and multiplayer. Every agent that is deployed with OCE comes with that built-in.

这是 Cooper 的仪表盘。我可以打开一个经过身份验证的会话，连接到 Cooper 团队的服务器。我们来看看。So here is Cooper's dashboard. And I can open up an authenticated session to Cooper's team server. Let's see.

好的。太棒了。现在，你可以在这里看到我刚刚进行的对话，还可以进一步引导对话，精确定义你希望如何创建这个功能。Okay. Perfect. Now, here you can see the conversation I just had, and you can further steer the conversation to precisely define how you want the feature to be created.

确保仪表板支持自定义配色方案。团队成员也可以登录同一会话，选择跟随操作或控制会话。Web 控制台。Make sure dashboard has custom color palettes. And teammates can also sign into the same session to either follow along or steer it. the web console.

好了。现在，库珀正在处理，我想弗雷德应该已经准备好了。我们来看看社区页面吧。All right. Now, while Cooper is working, I think Fred should be ready now. So let's check out the community page.

嗨，弗雷德。我们今天要发布什么？我可不是每天都有机会给彼得发消息。Hi, Fred. What are we launching today? It's not every day I get to message this Peter.

所以我想知道还能发送哪些其他消息。目前，Fred 正在查找 Linear 问题、浏览 Slack，然后利用这些上下文来回答这个问题。理想情况下，接下来的 10 秒内，我们要么开始工作，要么继续往下走。So I'm wondering what other messages I can send. Right now, what Fred is doing is he's looking for Linear issues, looking through Slack, and then using that context to answer this question. Ideally, in the next 10 seconds, or we're going to work, or we're going to go onwards.

我们有一个很大的 Linear 看板。好的，我们收到回复了。We have a big Linear board. Okay. We got a reply.

哎呀。我觉得我们的插件好像出了点问题。不过看起来我们确实拿到了 OpenClaw 的发布说明。Oops. I think we have some plug-in issues. But it does look like we got the OpenClaw release notes.

好了，今天我们有一个 Linear 工单。今天的工单是发布 OCE。所以，Fred，下次请做得更好一点。Well, we have a Linear issue for today. And the issue for today is launching OCE. So, Fred, please do better next time.

我来回顾一下。今天，我们正式推出 OpenClaw Enterprise。I'm going to recap. So, today we are launching OpenClaw Enterprise.

ClawLabs 与开源基础设施ClawLabs and Open-Source Infrastructure

这是一个与供应商无关的开源平台，用于管理和部署智能体。正如 Kubernetes 成为在云端部署容器的行业标准，我们也希望 OCE 能为智能体部署发挥同样的作用。OCE 最初由我们在 OpenAI 启动，并在开发早期捐赠给了 OpenClaw 基金会。This is a vendor-neutral open source platform for managing and deploying agents. Similar to how Kubernetes was the standard for deploying containers in the cloud, we envision OCE to do the same for agent deployments. We first started OCE at OpenAI, and we donated it to the OpenClaw Foundation early in its development.

在那里，我们与红帽和英伟达等合作伙伴继续构建它。今天，我们向全世界开放。来试试吧。There, we've continued building it with partners like Red Hat and NVIDIA. And today, we're opening it up to the world. Try it out.

请告诉我们如何改进，并帮助我们为智能体构建安全的基础。谢谢。Tell us how we can make it better. And help us build a secure foundation for agents. Thank you.

交给你了，彼得。嗨，谢谢你，凯文。我想回到我们最初谈到的地方。Back to you, Peter. Hey, thank you, Kevin. I want to come back to where we started.

OpenClaw 可以在底层使用 Codex，因为 Codex 是开源的。我们还扩大了 Codex 开源计划，让更多人能够使用 AI。此外，通过 Patch the Planet，我们正与 Trail of Bits 等团队合作，提升开源软件的安全性。OpenClaw can use Codex under the hood because Codex is open source. We also expanded the Codex open source program so more people have access to AI. And through Patch the Planet, we're working with groups like Trail of Bits to make open source more secure.

开源软件是关键基础设施，而我们的目标之一就是修复一整类漏洞。我们正是以这种方式用 OpenClaw 构建 OpenClaw。 如果你正在和自己的团队一起尝试共享代理工作流，尤其是规模较大的团队，我很乐意听听哪些做法有效，哪些行不通。Open source is critical infrastructure, and part of the ambition here is to fix a whole class of vulnerabilities. That's how we're building OpenClaw with OpenClaw today. If you're trying shared agent workflows with your own team, especially larger ones, I'd love to hear what's working and what isn't.

演讲结束后过来找我。谢谢。Come find me after the talk. Thank you.