OpenClaw周下载量超400万 架构师回应安全与商业化争议

OpenClaw周下载量超400万 架构师回应安全与商业化争议 OpenClaw基金会选择坚守非营利开源模式以保障全球信任，并通过Gateway架构与自动化测试协同推动Agent生态全面爆发。 本期访谈围绕 2026 年爆火的现象级开源项目 OpenClaw 展开。首席架构师 Vincent Koc 详细拆解了 OpenClaw 的常驻 Gateway 网关架构、心跳机制、内存管理与安全防护体系，并分享了团队如何利用内部 Agent 自动化处理海量 PR 和极速迭代。面对中美市场在采纳速度与企业管控上的文化差异，Vincent 强调 OpenClaw 坚守 501(c)(3) 非营利基金会定位，致力于保持开源与中立。最后，他预测异构 Agent 间的通信与协作协议（Agent-to-Agent）将成为未来 12 个月的核心突破点。 2026年初 OpenClaw 在全球掀起 AI Agent 狂潮，其首届旧金山 ClawCon 盛况空前，甚至在中国引发 Mac Mini 抢购断货与全员 KPI 驱动的落地热潮。 (0:00) 面对峰值约 1.3 万个 GitHub PR 积压，OpenClaw 构建内部 Agent 自动化审核系统，支撑起包含近百个代码库的庞大开源生态。 (5:00) 经历架构重构与稳定性修复后，OpenClaw 的 NPM 周下载量不降反升突破 400 万次，显示其实用价值已成功替代初期投机热炒。 (11:15) OpenClaw 核心采用具备常驻服务与心跳机制的 Gateway 网关架构，并引入 Vault 凭证加密、原生容器隔离与 Auto Mode 自动安全评估机制。 (23:00) 基金会坚持 501(c)(3) 非营利架构以扮演 Agent 领域的“瑞士”，并预测异构 Agent 间的跨平台通信与协议协同将成为未来一年的核心爆发点。 (50:00)

OpenClaw 首席架构师 Vincent Koc 指出，OpenClaw 正从早期的狂热试验转向以实用与稳定为核心的基础设施；通过非营利基金会的中立治理与原生架构演进，该项目致力于成为异构 Agent 生态的开放枢纽。 全球热潮分化与应用场景落地 (7:11) 2026 年初 OpenClaw 凭借“龙虾”吉祥物在全球引发关注，并在中美两地呈现截然不同的采纳路径。在西方，多数企业出于数据风险与安全审查采取谨慎防御态度；而中国市场则展现出极快落地速度，不仅出现员工被分配 OpenClaw 自动化 KPI 的现象，更一度引发 Apple Mac Mini 售罄热潮。Vincent Koc 认为，中国对开源替代品的渴求与高频迭代反馈，使之成为验证大规模部署的重要试验场。 在经历早期部署复杂与稳定性问题导致的卸载潮后，团队重构了底层架构。NPM 每周下载量不降反升，由炒作峰值的 250 万至 300 万次反弹并突破 400 万次（不含中国镜像源）。此外，OpenClaw 逐渐催生出以数字化协作者形态常驻于团队协作流中的新型办公模式，并在各大科技厂商推出衍生版本后步入各框架互通的“后 Claw 时代”。 Gateway 常驻架构与工程化自愈能力 (9:50) 区别于传统回合制、被动响应且运行环境固定的 Coding Agent，OpenClaw 采用常驻服务（Always-on Service）式的 Gateway 架构，依靠“心跳机制”（如每 10 至 30 分钟触发一次检测）在后台自主调度并执行周期任务。更核心的差异在于其具备修改自身代码（Self-modifying code）的能力，可根据新模型接口自主调整 Harness 套件以实现零天支持。 面对代码库扩展至近 100 个仓库以及峰值达 13000 个 GitHub PR 的挑战，团队搭建了由内部 Agent 主导的代码审查与测试“工厂”。测试框架支持在 20 秒内调起跨平台云端环境，由 Agent 自动录屏、截图并附于 PR 之中。在数据与安全层面，OpenClaw 将存储由纯文本文件重构为轻量级数据库（如 SQLite），并引入凭证 Vault 加密、容器化权限隔离，以及结合前沿大模型与本地端侧模型的混合“Auto Mode”意图安全审核机制。 反思多 Agent 复杂性与实用主义路线 (35:00) 面对业界堆砌 Manager、Planner 与 Sub-agents 等复杂分工的趋势，Vincent Koc 引用“康威定律”指出，许多企业只是机械地将既有人类组织架构映射进系统，继承了冗余包袱。OpenClaw 主张 Pragmatic（简约实用）哲学，认为在当前先进模型理解力大幅跃升的背景下，单 Agent（Single Agent）模式在 90% 的常规场景中比多 Agent 更高效且更节省上下文开销。 Vincent 坦言，长期记忆（Memory）在跨模型迁移与多租户权限隔离上面临极高技术复杂性，远未得到根本解决。因此 OpenClaw 的定位并非全包型黑盒，而是作为中立的协调者（Orchestrator），负责工具调用、环境调度与标准化规范制定，将深度的长行程自优化留给模型自身或垂直系统处理。 非营利中立治理与未来互通愿景 (50:00) 为避免被单一资本或商业利益绑架，OpenClaw 遵循创始人 Peter Steinberger 的设想，确立为美国 501(c)(3) 非营利基金会，致力于扮演 AI Agent 领域的“瑞士”。基金会依赖社区捐赠与生态赞助运作，不承担风投回报压力；核心工程团队仅维持 5 至 6 人规模，贯彻“Agent-First”原则，在新岗位设立前先由内部自动化 Agent 承担职能，成熟后再将工作流开源回哺社区。 针对未来 12 个月的发展，Vincent 预测技术演进的核心焦点将不再是炒作性的虚幻安全叙事，而是“异构 Agent 间通信与协作协议”（Agent-to-Agent Communication）。打通跨企业、跨模型平台的个人 Agent 与工作 Agent 对话，并依靠机制可解释性验证行为逻辑，将成为推动 Agentic AI 普及的关键命题。 OpenClaw 证明了开源个人 Agent 作为协作基础设施的可行性，但如何在消除跨模型持久记忆瓶颈、防御系统级越权风险的同时，推动异构 Agent 通信标准的统一，仍是行业迈向真正普惠所必须攻克的工程难题。

[20:00] Vincent Koc: It was OpenClaw's architecture that made it stand out... There was this promise of AI agents that would do things for you... and that promise just never came to life. Peter was just frustrated that it didn't happen, so he started building for himself. [50:00] Vincent Koc: We essentially want to become Switzerland for AI agents. [60:00] Vincent Koc: Peter's wishes were like, he thinks the only way the open clause survives and is fair to everyone is that it remains in a foundation and it's not controlled by one company or one country or one entity.

[0:00] OpenClaw在今年年初爆火之后在全球是掀起了龙虾热宣告正式迈进AIA agent 时代OpenClaw is the number one, is the most popularThe single most important release of software probably everLobster themed installation events have popped up around the globeEspecially in China众多科技企业纷纷推出Claw产品开启全名龙虾热潮然而很快人们发现没有想象中好用以及安全考虑又是出现了现在风潮那么个人Agent架构的下一步是什么?在上海WAIC现场我是采访到了OpenClaw Foundation的首席架构师Vincent Koch那么以下就是我与Vincent的视频播客对话嗨 Vincent,谢谢我们在WAIC2026在上海我们开始来介绍一下你自己的专门和分享一下你的专门的专门的专门我说,我说,我看这个,我看这个,我看这个,然后我说,哇,这个是在最后一年轮延的,然后那终于那终于我认为有点点关注我认为我认为在SanFrancisco是在 February where there was like a first ClawCon sort of event, and...我认为有数百个人呃 guest list but the event started at seven but if you arrived at five o'clock you were late and the queue was like around the building yeah i remember that one my business partner was there yeah it was crazy yeah she told me the queue is like a few blocks away already yeah it was and we hadn't seen anything like that sort of虽然 everyone is there everyone's listening in san francisco in quite some time it kind of had that magic to it like you felt like something is going to happen so for me like that was like the the peak sort of hype cycle and i think after then it just started taking off quite a bit but did you guys expect that you know the open club was got really really crazy in china no i was uh obviously still a maintainer but we're in the team on discord and i'm seeing like the auntiesI'm seeing rap artists on Twitter talking about Open ClawIt's just crazy and I'm just like everywhereIt's just going completely viralChina but also outside of ChinaIt's about China especiallyAnd a lot of people like this is fakeAnd I was like no it's not fakeIt's definitely realTell us how it looks like to work with Peter SteinbergerThe creator of Open ClawI think there's only a few people I've had the chance to work withThat have like a very special way of seeing the worldthe possibilities of what you can doI think his attitude to life aroundyou know make sure you put fun and energy into everythingthat you do definitely comes acrossin the work like you look at the fun little lobsterand all of this soI think for me has been not onlyjust about you know creating what's new what'son the bleeding edge but alsolike that process of discovering somethingexciting has to be fun you haveto think like a child you have to kind of like pushyourself and you're thinking a little bit soI think for me it's no很感激感到你说你加入了一个社会的社会在 December的年纪,我们只有一半年后,我们看到 OpenClaw的组织成为一个小传统的社会,一个社会的社会,一个社会的社会。什么是一些的大挑战你对你有多多多的挑战?我觉得,我认为,一切的挑战是,一切的大量是,一切的大量是,一切的大量变得更多。很难,不去管理,但去你建造一些 software那是 designed for developers or tinkerers and now suddenly you have enterprise companies deploying this or trying [5:00] to use it to build their businesses, right? So then you start getting people screaming, like, is it safe? Or how do we deploy this at scale? Why is this using so many tokens and things like that? So you start realizing there's other challenges and everyone's like expecting you to fix this. when actually it's just an open source product and we need to sort of like educate people like, hey, look, we're going to do our best. We're going to try and help and solve this. But the frustrations that we got and the challenges that we faced were ended up becoming the things that we started building for as well. So for example... we ended up with, I don't know, at one point we had like 13,000 PRs open, pull requests, like changes to OpenClaw. And we developed our own review using AI agents to like manage this workload. So then we started building... the other agentic tools to help us with the scaling problems that we're having. And I think that was like the unlock for us. And that was like the more exciting part. So we had a lot of problems because the existing tools out there just weren't designed for our scale. Yeah. But now it's getting better. Yeah. Like we've had to essentially build things. entire infrastructure. So if you look on the OpenClaw website, there's a little link called Ecosystem. We have over, I think, close to 100 repos. So a lot of people know OpenClaw, the agent, but also we have different libraries. We have different automation tools, testing infrastructure, and we've had to build all of these. And there's a joke that each one of these tools are basically like a startup in themselves that we manage and enables us to build OpenClaw at scale. If you have to describe OpenClaw in a few sentences today, what would they be? the OpenClaw Foundation is not, I think it's more than just this AI agent or personal AI agent. For us, it's like how do we empower the whole AI ecosystem, agentic use, be open and accessible to everyone. That could be for enterprises, that could be for small businesses, it could be for moms and dads, it could be people that want to run local models at home as well. So For us, the mission is to really empower humanity with AI. Did that mission change at all during the past six months? I think that was pretty clear. You know, when it started growing legs of its own, the foundation was already being set up in the background. Obviously, it's been announced recently. So the mission statement was, I think, as a... collective was somewhat already there. And the idea has always been like, we want to keep Peter's wishes were to like, you know, he wants to keep it an open and independent product and put it into a foundation. And for the same reasons, it's like, we want to make sure it's accessible to everyone. It's continues to remain open source, continues to remain in the hands of the community. which was very important. Before we dive into the technical questions, let's talk about the China sensation. In the first quarter, as you described, the OpenClaw got extremely popular in China. I would say even more popular than Silicon Valley. And there were long queues of people lining up to install OpenClaw. Apple Mac Media was totally sold out back then. So why in China though? I think there's a mindset thing that I've noticed, especially in China. It's like, us, it's going to help us Let's embrace it. I think the... The Western sort of tech ecosystem was like, wow, this is scary. It can do so many things. Let's block this and let's understand what's happening. And we saw this sort of behavior like back in 2023, 2024 with like ChatGPT as well, where enterprises were like, we have to block this AI because like, it's too scary. It's too dangerous. And then eventually they opened up and adopted it. But I think for China has been very much like, I want to adopt it at speed. I think, and some of the stories I was hearing was like, employees were given like KPIs, like you need to automate this many things with OpenClaw. Whereas if you install open core in some enterprises in other markets, like you get in trouble because it's like security risk. Right. So I think the mindset is very different around like I want to experiment, learn and build fast. But I think it's just a just a difference in the technology culture and sort of the speed in which China has been moving, especially with open source. Is it possible? because, you know, like Codex or Cloud Code is not that popular in China. That's actually a good point. Yeah, it's not something I really actually thought about because the market penetration of these other products, you know, they can't be really used easily or used at all within the China ecosystem. So I think having something that's open source and available that people can adopt could have been obviously one of the key catalysts as well for this. For people who are not familiar with OpenClaw, can you tell them the difference between OpenClaw and the coding agent? So a lot of these coding agents are starting to become like OpenClaw as well. So it's kind of interesting. But if we rewind back to the start of the year, the coding agents were very much like, "Hey, I have a terminal or an app, I want to build [10:00] I don't want to build a website or something. So you have a conversation and it goes off and builds it. The difference between now, like OpenClaw, OpenClaw could, Yes, build your website. But the whole idea was that it acts as your personal agent. So it would learn. It will have a memory. It would start to build skills. It would go and do things for you. But the beauty is that, like, those things could also be in a secret code. So... oh, can you buy my groceries from this website? There's an API, it finds it, so then it creates the code to do it. Or for example, let's just say tomorrow one of the labs releases a new model and there's no support, so you ask the open call, hey, can you add this feature? And it can change its own code and adapt itself and add that feature for you. So I think the difference between the coding agents is like it's creating code and it's blocked, like the harness or the tool, whereas the open call agent can change its own code. The agent can... Modify itself, which I think is the difference here. Smarter. Yeah. But so many people had to uninstall OpenClaw after a while. Setting number one, it's not so easy to use. Number two, security concerns. How do you see the cooldown of the interest? So one of the two things with that that's actually interesting is that there was definitely a cool down of the hype cycle. Like the craziness sort of died out a little bit, which was kind of good. But the second thing is like, we did mess it up a little bit where... We were trying to build this thing really fast using agents, really trying to push the frontier here. And we learned some good lessons on how not to build software with agents. So we did make some mistakes and we did lose a lot of users because it wasn't as stable. But ironically, we overcame those challenges a couple months ago. And if you look at our download numbers now, they're actually better than they ever were. So before the craziness of the hype, I think we were hitting like two and a half, three million downloads a week on NPM. So obviously, there's Chinese Mirror websites. other sources. So download numbers are probably way larger, but just based on NPM. And a few weeks ago, we had like four point something million downloads a week. So yes, it did sort of die out a little bit, the hype, but actually the adoption is actually scaling even more now. So it's kind of interesting to see. So from that perspective, I'm like really excited to see that like, you know, users are responding to the stability and the features that we're releasing into the market. Can you give us a few examples? What are some of the really good use examples that you see across different industries? So one of the things here is that like, I think we talk about industry specific use cases. I like to think of it as like, it's just like having another employee or internal personal assistant. So, Yes, it can be vertical specific use cases, but realistically, it's like a case of if a model is good at doing some code or has understand some tasks really well, and it has the tools or it can create the tools. I think it can do a lot of things now. So it can now browse websites. It can have a phone number. It can connect a messaging app. So really, the possibilities are kind of endless, but depends on the task, right? So maybe some design tasks are really hard or video editing might be hard. your agent can then use it. For example... our agent creates soundtracks for our team every few days and it makes it kind of fun for us, right? Whereas AI for music was kind of not so good before, but now it's good. So then we can give that tool to our agent. So for us, it's like less about use case. And I think it's more about, you kind of discovering what's possible. One thing I do see that's interesting inside of enterprises and large sort of small to medium sized enterprises is also like employees having open clause in their kind of day to day job or projects having them. And I think that's changing the way employees work with agents. Because previously, if I was an employee, like an engineer, I would use my own AI coding assistants, and I would work and then share that with the team. Whereas now the it's not a coding assistant anymore, it's an agent, but it lives with the right so it can talk to my colleagues and my colleagues can talk to my agent and it has like a different kind of yeah so i think That's creating a different dynamic for us in terms of workplaces and how we operate. And I think some organizations here in China and globally are starting to experiment with this, and some are kind of pushing the boundaries around that. Obviously, we internally use agents and part of our workflow and talk to agents on a regular basis. But I think that's where it starts to get really interesting because it's less about use case, but it's changing the way we work, actually. Back to the first quarter, Many Chinese companies released agent products on the top of the OpenClaw framework, including Tencent QClaw, Volcano Engine ArcClaw, Alibaba, Wukong, Baidu, Duclaw, etc. So many. Do you think that's the right approach for the 2C consumers? I mean, like these big companies, they actually launch their own claw products. Yeah. So a couple of things here, especially for us, like we were just pushing this. open source product. China was running at fast speed. They wanted to get it to the market. They wanted to get it in [15:00] of users and their customers really quickly. So they developed on top of it, right? We didn't have the capability to do it at the time. Now, fast forward a little bit. Do we want to go straight to consumers with OpenClaw? Like we kind of do that at the moment, but realistically, like, at the end of the day, we just want to empower people the ecosystem, right? And if that means that other companies are building on top of OpenCore and getting that in the hands of users and empowering their users, I think that's a great outcome for us. But also at the same time, The other key thing here is that if we wanted to build... infrastructure and servers and hosted for people it kind of feels like it's taking us away from like what we're trying to achieve Which is actually building this software that works really well and helping the ecosystem sort of get that in the hands of everyone else That might change in the future, but that's kind of like where we see things at the moment so you would have seen also recently in the US we've had a Microsoft Scout, which is Microsoft's version of OpenClaw that's sort of out there. And other companies are going to continue to do this. And for us, it's like, how do we support the ecosystem where they want to build on top of it? That could be a large technology company, but it could also be a small enterprise that wants to do something internally as well. A lot of the products I mentioned, unfortunately, did not work out pretty well. What's the reason, in your opinion? I haven't had a chance to actually experiment with them. So a bit hard for me to sort of answer that question. But I would say that possibly... I don't know if it was features or if it was cost or what those things were, but with anything, it's probably worth sort of digging into. But yeah, I haven't had a chance to really dig into those products in detail. Now, China tech companies later developed their own agent products. For example, Tencent WorkBuddy, now very popular, but it's not directly built on OpenClaw. However, it is fully compatible with OpenClaw skills and frameworks. How does OpenClaw actually mean to this? latest rounds of agent products. So for us, I kind of see this is like this post-Claw era. I believe that OpenClaw really ushered this next generation of agentic frameworks into the ecosystem. It kind of created this explosion of model labs and token usage. So at the end of the day, in terms of what we're trying to achieve, I think it's good to have... other solutions out there in the market. And that kind of enables like a really good ecosystem. So I think it's a positive thing for the consumer and the users and the ecosystem as a whole. And how do you continue to work with Big Corp? partnerships. That one's a bit more interesting, but I think currently our model is very much the case of like on a case by case basis. But realistically, we try to create one solution that fits most people. And that's where, you know, we have to bring a lot of differences of opinion, we have to bring a lot of different technological ideas into the same place, and get everyone to sort of agree on like, what's the right way to do something. with like, okay, this is how we should solve this problem and this is how it should be designed. And by doing so, we kind of help propelling the industry forward as well. I know you've been traveling a lot these days, Japan, Singapore, and of course come to China a lot. So when you do the partnership with the big corps in different countries, are the partnerships different? Are the requests, demand different in your observation? I would say less so partnerships, but the companies that I've spoken to, and this goes beyond just like talking partnerships specifically, but just I like to spend a lot of time with developers, with enterprises, with the ecosystem as a whole. I would say sometimes the challenges may seem different, but I noticed that as a whole, they're all pretty much the same, right? For example, if you're in a country that has a large financial sector, you might be worried about what happens if your financial sector starts adopting something like OpenClaw. Is that going to be a risk? Is there some security concerns? So a lot of the challenges a lot of organizations have. and governments and policymakers and other people talking about are going to be relatively the same thing but depending on where they are with their digital maturity it's probably going to change the the lens of of the sort of the complexities of the questions but majority of it sort of revolves around um you know how do we get value out of this how do we uh manage costs um how do we make sure this is safe how do we scale this properly like i think this is not an open-claw specific thing, but just AI in general when you speak to anyone. So I think any enterprise right now, or any large organization, or anyone who's technical is thinking through these things when it comes to the AI strategy. What does the China market bring to the open-claw community? Is there a lot of contributors, a lot of users, big core partnerships, or continuous feedback? I think it's a combination of all of those. The benefit of the Chinese [20:00] This market is obviously, it's been faster than adopting OpenClaw faster than anywhere else. So for us, it's like, It's kind of a window into what's going to happen next in a way as well. But also there's large enterprises here that are using it, right? I know of many enterprises here that have like every employee is using OpenClaw. The projects are running on OpenClaw. And it's just interesting to see... the dynamics and to work with those companies and just understand like, what are some of their challenges? What are their learnings? How can we share some of those? So I think that's definitely the case. And, and we have some maintainers on the team that actually are situated here in China as well. So we have that kind of mix. global mix when it comes to our team and our roadmap and the features that we work on and also the developers that we listen to as well. And next, I want to spend a few minutes to walk through some of the architecture and the really important technology questions. It was OpenClause architecture that made it stand out. The gateway, agent runtime sessions, memory scale, plug-in models, messaging channels, and connected devices work together in order to have the agent complete a task. work what is still the challenge and then what needs to be improving uh to make open claw work better i think what what was missing and you can ask anyone this is like there was this promise of ai agents that would do things for you that you could talk to and it would understand you and it can do things to an extent and that promise just never came to life right you know yeah we had coding agents but like a coding agent is not the same as like my personal ai everyone wants personal Peter was just like frustrated that it didn't happen. So he started building for himself and eventually a community formed around it. And then it became open cause, you know, now. So I think, that the fundamental needs of the users with like what people wanted were just like missing. So I think that's the critical difference, I think, that kind of came about. And I think maybe it was a combination of timing, The community, maybe a different number of factors that sort of help bring that to life. And when we think about some of the challenges now, I think for us specifically, we know that the models are getting better. Open source frontier models are getting better. So knowing that the models are going to get better, what are some of the other challenges that we need to face? I think a lot of it comes down to things like scaling within an enterprise. There's going to be some specific technological requirements. in OpenClaw before we used to write everything into files. So all your session... conversations. That's now put into a database that's in a little file. What that means is that if you're an enterprise building this at scale, you can replace that file with a database system, you can take backups, it makes this a lot more scalable, right? And easier to deploy. So yes, it improves the performance for consumer use for like personal use. But also at the same time, it means that for enterprises like, or technology companies, like you mentioned, that are like building on top of these makes it really easy to build on top of OpenClaw. So a lot of what we're focused on functionality around how do we build this thing in a way that, that enables more people to build on top of OpenClaw for whatever their use cases might be or their business challenges might be. Let's talk some of the tech details. So first of all, Gateway. Gateway is the central part of OpenClaw's design. It appears to be the central nerve system of OpenClaw connecting messaging channels, models, memory, and tools. Why did OpenClaw choose this always-on Gateway architecture? And what can it do that a conventional agent framework cannot? So the difference with why did this happen, you'd have to ask Peter this question. I think it was his design or maybe his agent came up with this design or back and forth. They came up with this design. I'm sure there's a story. I just can't remember off the top of my head. But I would say the difference at the time was that existing coding agents were basically turn based. Right. You had a thread. And they were turn-based. The difference with this sort of gateway concept is that, like, it was always on. So it's like working like a service, essentially, on your machine. And it would connect to... the messaging channels, and then that's where the concept of the heartbeat came in. So it's like every 30 minutes, every 10 minutes, it's going to trigger a conversation chain. It's going to check if it needs to do anything. If it needs to do anything, it'll go off and do that thing. So it kind of created this like, human-like sort of like, I'm going to go and do things in the background for you and have a way to like schedule tasks and with the agent and do things. I think that's what started giving this OpenClaw a little bit of its magic as well. And also the gateway may hold credentials and execute actions across the user system, right? It is also OpenClaw's biggest security boundary. So what has OpenClaw community done to enforce the protection list? [25:00] Yeah, so a few things there. When it comes to credential storage, like we've implemented some security features where the credentials are not stored plain text like they're in a sort of vault sort of concept. But essentially, one of the key things that we are working towards is like containerization, right? So, You can run open call in a container now, so you can say that open call, that agent, It can only have access to these files, this internet connection, things like that. But we're taking that a step further and we're sort of like working with industry partners, to like bring that as like a more native feature without saying too much. But, um, where if you say, It can only access these files there's a way that the operating system and OpenClock can kind of know 100% it's only accessing these files. And if it needs to, it goes back to the user itself. and then ask for permission, right? We're creating these sort of security boundaries outside of OpenClaw as well and working within the street to make that a reality. So that's definitely in place. And parts of this already exist now, but parts of this will roll out as other technologies around security also happen. Because you also got to remember, right, like, OpenCore is like one piece of the puzzle. You have the models, you have the operating system, you have the, you know, we have to work across industry to sort of, like, solve this problem. It's not, like, exclusively us. but then you would have the same security concerns with any other coding agent. If you gave a coding agent, like, full admin access to your machine, it can also delete files and go to the internet and do all these things. So I think a lot of times... A lot of people direct towards us. we're more than happy to sort of help solve the problem as well but i would say it's like still a agent-specific sort of challenge, but we are solving this more from like a containerization perspective. So working out ways that you can kind of put it in a box safely and say, hey, you can only do these things. And when a user accepts it or doesn't accept it, like it's very clear and we know that that's actually happened. I think you briefly mentioned this before, but what actions should an agent be permitted to take on autonomously? And what kind of actions ask for approval. I think that's really hard to answer because like it's going to come down to what your intentions are, what use case you're working with. You know, you might decide, hey, deleting files on my machine is dangerous. But you might be working on another project where it's just a test machine, whatever, and you're like, delete them, I don't care. but like they're not the same thing, but they are like your intentions of like what you're trying to achieve is like going to be quite important. So if you look at other coding agents, they've rolled out things like automatic, uh, sort of approval using large language model to understand the conversation thread and do this. So we've implemented a similar feature called auto mode. So instead of having on, We have like this in between where an agent will look at your conversation and then decide based on your intent, is this safe or unsafe and then ask for approval. So we've sort of implemented that at the open core level, which means that if you're using an open source model or some other model, you still get this feature. Now, the other benefit with this is, well, I might be using an open source model that might not have maybe Some of the security guardrails, the auto mode thing could come from like a frontier lab model. So I could pay for that access and just have like that safekeeping what my agent is doing. But then I could run all my inference on, say, like a local model. So then I kind of get a hybrid. of these two worlds as well. So it looks like a lot can be done actually through different layers. And you talked about models and we're now experiencing another round of SOTA model racing in Silicon Valley. And in China. Also in China. That's right. How much does model update actually impact the agent framework? And how much does that capability comes from OpenCloud's agent harness memory tools or executing environment? I think it's a combination of both. Yeah. Yeah. I would also say that some models have also been trained, post-trained on OpenClaw. I know of some open source models, especially some of the ones in China have definitely been post-trained on OpenClaw, which means that those models then perform better because they understand, like the model just knows that it's inside of OpenClaw and the memory and the tools and things like that. So there's definitely like a combination of the two. But we also know that some models... don't work that well. So which tells us that like, the model needs to be at a certain level of intelligence or size. And we're currently working through a process of running a benchmark and evaluation that we're going to make public about 120 tasks, 100 tasks. We're working with some research teams and some labs. And we're going to release that over the next few weeks, hopefully it's been taking some time. And that's going to help us understand where the gaps are And we're going to kind of release that to the market. But then we're going to work on, okay, where we know that the certain models or certain tasks are failing correctly. [30:00] What can we do to improve that? improve the context or can we change something with the memory when it's a smaller model to like make sure that that performance works really well so i think there's a combination of the model but also how the model works in the harness and we were seeing this with the front a model labs they're all releasing harnesses as well to complement their model and I think it's the combination of the two right because those harnesses come with tools and memory and other things and those models have been trained to explicitly use those tools and I think that's what making them powerful. And OpenClaw supports models from different providers. Correct. How can you guarantee the model independence? In what way? So that you trade actually models fairly, not to favor certain models. Yeah, so we support any model release that comes out. We try our best to get zero-day support. We work with all the different model providers and labs like that. And yeah, we don't specifically push one company or another company. We make sure that feature parity exists as well, and we do testing as well. So I think it just comes down to... capability. We also kind of built some features in there which allows model companies when they've they don't need to run an update on our site. They can push some changes on their end. So that makes it sort of fair and a little bit more easier for everyone else. But I think the benchmarking process and the tuning is going to help us, like especially with some of the smaller models, make sure that there's like a better experience for users as well in OpenClaw. Are smaller models important for the agent work? I'd say more for the local, like the local model, open source sort of community. specific tasks, people are obviously concerned around token usage. So for us, it's like, how do we make sure that you get some experience? It might not be the best, but like you might be able to carry out some autonomous tasks with OpenClaw. So for us, it's like a really good learning experience to sort of figure out how do we then architect things parts of OpenClaw to then be fair for all of these models, right? Small, large, medium. So that's kind of part of the process as well. Memory. Persistent memory is one of the features that can make a personal agent genuinely useful and personalized. But it can also introduce errors, privacy concerns, and unwanted assumptions. How do you balance these two sides? I would say the interesting thing with memory is like a lot of people might say this like a solved problem. In my mind, this is like 100% not a solved problem. I think memory is like one of these things where it's like really complex space. Like you've even answered it, right? Like for example, let's just say I'm using one model and the memory files get created and then I switch to another model. How do we know that that model understands? There's all these like complicated factors, right? Other issues arise when you use like multi-tenancy. Like what happens if you're sharing one claw in a team? which memories can be shared with which people. And that's why we, for now, basically say with multi-tenancy, it's like that claw, assume everything is... interchangeable. There's ways we're working on solving this sort of multi-tenancy thing at the moment but you know that that's why it's like a not a solved problem and we're actively working on this and looking at different ways of solving it that's probably like one of the more interesting harder problem spaces right now for sure yeah because i think that's what gives it the the capability to do what it needs to do. Right. Yeah. And also long horizon tasks are pretty challenging, but necessary. Correct. For the next chapter. What's your approach? For long horizon tasks, I believe a lot of the newer SOTA models are like getting better at long horizon tasks in general. We're also seeing that industry is building just for context, like prior to my work on OpenClaw, I was working on self-optimizing agents. improve themselves for like a particular long horizon task. Yeah. So I think, The frameworks and approaches are going to come out from industry around that. And I personally believe they're just going to become capabilities that we can just give to agents as like tools or some sort of features that we can build into it and let it manage it. So, I think the best way to describe it is that OpenClaw could act as an orchestrator and keep on top of your work and do all of that and use different models to orchestrate that work. But I don't see us solving long horizon as a problem, right? Whereas different model companies and systems might, and then we find a way to bring that into the OpenClaw system. And for the multi-agent design within OpenClaw, when is it useful to route work to isolated agents or specialist agents? additional cost latency as well as coordination problems. I think it's less about cost and latency but more of a case of like context like most nine times out of ten for most people one single agent is like more than to four. [35:00] to like handle a lot of complicated... stuff where it happens it's like oh there's this other very particular task i want to run every couple of days do a certain thing it has its own system prompt once you start like carving that out then you're like okay I can just make that a separate agent. The best way to think about it is like, I have one person, one employee working for me. Now their workload is a lot. And I know certain workloads is like very repeatable. Okay. How do I take that work off them and then give that to someone else? So you kind of think more like I'm managing someone. How do I break this workload up between people? It's like a better sort of mindset to sort of apply to this when you're thinking about this. And we are seeing some agent systems are moving towards a more complex structure involving managers, planners, sub-agents and nested team of agents. However, OpenCore actually continues to keep it simple. Yeah. so-called pragmatic approach. Why is that? So in engineering, we have this concept called Conway's Law, which basically means that most organizations... their software code will map their internal organization structure. What do I mean by this? Let's say my marketing team had a growth team, and like a retention team When I build the code, I might have a separate like promotion system and a separate like retention data system, right? Just because that's how my teams work. So I want one team to manage this code, manage another. I think, Complex organizations are mirroring their internal organization design into their agents because they have a certain organizational structure, they create agents in that way. it doesn't mean that that's the optimum way to solve things, right? So I think we're inherently taking on the past challenges and not thinking from a clean piece of paper, like what would be the ideal way and just building from the ground up. And I think this is why some startups and some newer organizations are able to build agent first and work at rapid speed versus some other organizations that have to go through a full change management process and really understand what does this mean. like transformation ai transformation like a like a global scale and we were having a conversation and she was saying to me this is just me sharing some personal notes by the way it was like you know ctos have all these resources they've all this like agent designs and frameworks but if i'm a chief operating officer like what does the future of the workplace look like like what does the team structure and the agents like what does that all look like and i think those are like the conversations that people need to have because they're just kind of taking what they have and then just putting into an agent expecting it to magically work because like i think where some of and just on. So people need to have the right expectation of the agent. Or the way to get that expectation is through experimentation, right? Like start with a blank piece of paper. I'm going to redesign this team or this problem. Let's build some agents. Let's build something in a day. Let's get it working. It's not going to work perfectly. What do we learn? Let's iterate on that. Whereas instead of taking small, small pieces and trying to solve this problem and make it perfect and i think it's more a case of like how do we increase the maturity of of like this agentic design and the way i keep saying to people is like 2026 way of agentic engineering is very different to how AI and agents were being developed today. prior to that. And I think a lot of organizations and individual developers and product people just haven't quite crossed... to the other side, yeah, and realize that actually, no, the models, like the SOTA models we're speaking about, have really good understanding of this problem and you can just talk to it and like work with it and build the agent together right with the agent and get it to a point and then you can learn from that and you can apply the learnings instead it's like oh i'm going to build all these little skills and then see what happens it's just you know it's like a i think you can just like leapfrog basically and learn from that experience yeah as you said there's like a surge of skills and plugins and and core framework within this open source community while the ecosystem is expanding. Do you see OpenClaw actually will keep that small core or it will slowly expand to do more like third-party ecosystems by itself? I think it's hard to say, but like we're definitely looking at ways that we can support the wider ecosystem in different ways. And some of our plugin architecture has definitely been changing shifts. It's not like we have a new concept like MCP or skills. I don't think we've had one of those in a while. But it's more... How do we build on top of that? Like an example the other day that we're talking about internally, what happens if one skill is dependent on another skill? Right now the skill spec doesn't accept that. There's no such concept of this in the industry. So we're like, do we adopt this and build this? [40:00] force the industry to do it. I think that's where we would work with the ecosystem and figure out where some of the challenges are and how do we think about what's next. I think you mentioned that Peter and other core contributors were wondering why agents' aha moment hasn't come yet. So is it a technical bottleneck? There's cost. There's how reliable is the model. There's all kinds of factors around this, right? And again, going back to my point, it's not just an open-clause specific issue. It's also the models and other things that encompass this ecosystem, right? And you might have all of the companies that you're using, like the banks, for example. Do the banks allow your agents to come in and transact and move your money around and do this thing? So I think there's a lot of challenges with like… the services that your agent wants to access and how do they access it. And so I think like the whole industry is like sort of growing up as well. So I think we'll start to see that. becoming more reality as services are like allow agents and they're safe and the models work and you know i think it's going to take a little bit of time but we are definitely seeing a lot of improvements the biggest unlock for me i've seen is the use of computer use models where you know like any one of the agents some agents support it and some of the harnesses we're also looking at like some open source computer use um stuff on our harness so that you can use other models as well and the idea is that like it can just take over a web browser right start clicking things and doing with your permission, which might be like a way to sort of unlock some of the challenges people are facing as well. After the first wave of the excitement in the first quarter, we talked about the cool down. Did the team, did the open cloud community actually learn something from the cool down? I think for us, it was like we just need to make sure that we build reliable software that works and works really well. But trying to do this in a way where agents are the main engineers and you're engineering the agents to do this. is very, very hard. And we had to spend a lot of time building the tools for this, right? So like one of them was building extensive testing tools. So right now, any one of our agents, any one of the maintainers can, their agents can spin up any operating system on any type of hardware on most clouds in like 20 seconds. And they can go inside, take screenshots, do video recording, click through. And that's how the agent's able to test and validate the changes and make sure that it's working and allow us as maintainers to go in and see for ourselves and see proof. So if I say, hey, attach proof to the PR, we'll see screenshot before, after video recording. But hey, I want to access the machine. It will say, okay, here, boom, I opened a web browser. I've logged you in. You can now take over this test machine. So we had to build a lot of these things that didn't exist. And I think that's one of our learnings was actually like to make agentic engineering work at scale. We call it a factory. The factory needs all the other machines to work, right? You can't just have one conveyor belt. expecting it to work you need the quality assurance you need the lights you need all these other things to make a factory run and the water electricity right so for us it was like we had to build all this other mechanics in that we're missing and we are now seeing the adoption of individual developers and doing like personal ai we're also seeing the enterprise side that they're adopting a lot of open open core framework as well so um to see and to be this two approach which one sure. One is feeding the other, right? So as consumers adopt it, some of those consumers work in enterprises and they're like, okay, how do I bring this into a business setting? But then as enterprises adopt it, it's making the product better at scale for different use cases, which makes it better for consumers to adopt it because it's more reliable because enterprises can use it. So it's kind of creating this really interesting flywheel for us. perspective, right? Sales, promotion, pricing, and the feature sets changes. Because we're not technically selling the product, It really comes down to feature differentiation. And because OpenClaw is quite extendable, allows enterprises to really make it and shape it into what they need. So I think we're kind of in this like really good sweet spot, which kind of fits most people's needs currently. But the challenge is obviously working with industry and working with the ecosystem to like, figure out how to balance everyone's needs properly. You guys actually launched the mobile app for iOS and Android. Why did you do that? So we actually had the apps built for a long time. I think it was still in the code base in like January, February. We hadn't registered an entity. So the foundation got stood up as an entity. And because it was stood up, we could finally go to Apple and Google and go, hey, Can you give us a developer account and we'll publish it, right? We didn't want to publish it under some... contributor's name. We wanted to do it officially. So we had them and we wanted to release them. I think some people want to [45:00] first party experience. and some people really enjoy that first party experience and some people might not so I think for us it's just giving people optionality on how they want to use it so for us it's not any extra work because we've already done the work building these apps it's just more a case of getting into the hands of people and I saw some less positive feedback from some of the consumers, what is the biggest challenge there? I think the thing with people is, especially on the internet, is like very quick to criticize. I think for us has been we develop at speed. And the beauty with all that criticism was like, great, we now have tons of feedback that we can incorporate. So since then we've made a number of releases, and Even within the space of two weeks since we've released the apps, if you look at where they are now versus where they were before, they look totally different. It's almost a completely different product. The beauty with open source is that other people could contribute. Luckily, we had people making contributions, which was good. But then a lot of designers were, for example, sending us nice designs that they created with AI possibly going implement this and you're like okay can you do a pull request can you like help us get this in they're like I don't know how to do that so a lot of people are quite quick to like make comments about software but I think people forget that it's like The OpenCore Foundation main engineering team is like five or six people. Yeah, we have other contributors and people working across the ecosystem as well. But we're not some 200-person startup, a large enterprise, but they're kind of expecting us. They're like, quick, quick, quick, can you do this? But we took on all the feedback and we made some quick changes. And we'd rather get the product into the hands of people, get feedback. then to try and make it perfect and spend months and it's not what people want. I think this is what I was saying earlier about how enterprises need to think about agentic engineering. It's like we can move quickly. So if you can move quickly, you're better to just get it out there, get some feedback and then iterate quickly. If people made bad comments and then we did nothing about it, then they'll be like, okay, that's bad. But... we were able to sort of like really change and shift and make a lot of changes. The same time, the websites got redesigned, our documentation got redesigned, like a whole bunch of stuff got redesigned. We're still hiring for product designers. as well onto the team. But yeah. Yeah, and to be honest, I was a little bit surprised that you guys launched the app because I thought that's what a startup would do. Yeah, I mean, the testing tools have been pretty good. We have testing tools and they'll go and launch it on phones and take screenshots and videos and maintainers. We just kind of compare like, yeah, this video looks good. It's like approved. So it does all the work, right? We still need to drive it. but we can do a lot more than we could before. The apps aren't super complex, right? Like, A lot of the features already exist in the web UI, so it's more a case of the same feature set in Android and iOS. The other thing we've also done is we're not using some framework. We've written it in fully native Android code, fully native iOS code. So we're not having to battle... complex code because the AI agents can write native code really easily. So I think there's some architectural decisions we made that make it really easy to maintain as well. Do you guys see Manus or Jaspark as a competitor? I haven't really used their products lately, so... Hot answer. But I don't know, for us, it's like I don't particularly see anyone as competitors in the industry. At the end of the day, we're trying to... propel the agent industry forward in this case. And I would say like our adoption and growth is like, just quite large and we're building more like a harness layer, right? We're building this sort of personal AI agent layer and now for enterprises as well. And it's very different from some of these other products and services. Some of these are closed source, closed source, SaaS products, which is very different from having like an open source product that you can take internally into your organization and build a business on top of as well. I have to mention another competitor of yours, although you don't want to call anything competitor, which is a Hermes agent. A lot of developers, they express a preference for Hermes agent, describing it as leaner or more coherent for certain user cases. Did OpenClaw team actually learn anything from their memory, they're very different and divergent from some of our features and the way our features work. So they kind of give you this like very... out of the box sort of experience. I would say some of the stuff they've done around memory is very different and kind of interesting. Some of the onboarding is kind of very interesting as well. But I would say we're quite fundamentally different in terms of products. Although on the surface, yes, They're essentially personal agents. You can connect them up to most messaging channels. But some of the internal mechanics around it are kind of very different as well. And do you see more open source agent frameworks coming out to the market now? And also, do you expect the... the agent framework eventually to consolidate around a few dummy [50:00] and like platforms, which is, you know, similar to the operating systems. Yeah. I mean, you could argue, yeah, maybe we will see some consolidation. But like if we looked at any of the AI tooling out there in the market right now, like none of that's happened. We have like hundreds and thousands of like AI software companies. Right. So hard to say what the future is going to look like. But yeah, we're going to see more and more agent frameworks. I think it's just going to accelerate. Everyone's going to create like maybe there's going to be vertical specific ones or industry specific ones or whatever the needs are. be niche products obviously to tap into that specific market but at the like larger scale there's only going to be room for like a small handful maybe but we'll see like it's really hard to say like this industry literally changes by the hour like i can open up my phone and look at my newsfeed i'm sure some new release of something crazy is happening right now right like and but i think that's also exciting that there's a there's a lot of opportunity there for anyone who's building, but also for us as well, to be able to sort of support the community as well, which I think is a really good thing. Next, let's talk about the foundation. Can you explain to us how the OpenClaw Foundation operates in practice? Essentially, it's a 501c. It's a non-profit registered out of the US. Because of that, we obviously don't make revenue. If we do make any sort of revenue, it's obviously going back into the community, back into the product and back into the ecosystem. And our focus on our mission is essentially what I've been sort of focusing on around like making sure personal AI and AI agents is like available and accessible to everyone and empowering sort of everyone. The operating model. So we kind of work like a startup when it comes to our engineering. But obviously being a non-profit organization, we have to sort of work in a certain way as well. So we have governance and all these structures and stuff like that. But predominantly, we get donations from the community. We get donations from the tech ecosystem that then drives innovation. a lot of the hiring and different product initiatives that we're doing at the moment. So you would have seen we recently made an announcement of some of the partners that have and some of the people in the industry that have been really grateful and sort of supporting our mission. And that's usually the sort of the driving force behind it. And it's very different from, say, like a venture-backed business that has to scale up to a certain number of multiple and hit a certain revenue number and things like that. Ours is mostly focused on what impact are we doing into the ecosystem and what impact are we doing to enterprises, personal users, the whole AI community as a whole, which means that our mission is very different. We essentially want to become Switzerland for AI agents. from the community or the foundation where you actually need a lot of talents. It's hard to say what's enough or not enough. You could argue that there's unlimited work and it's never enough. But I would say that we're very grateful for everyone's support. We get support from people giving us small, even $5 donations on GitHub down to supporting us with like, hey, we want to give you some tools to use for certain things. So I think support comes and I think we're very grateful for the community and different partners out there and everyone that's kind of been supporting our mission. Do you have the guidance of what kind of money you can absolutely take? I think internally we do have some you know we don't just take you know anything from anyone but that's something I'll have to check and come back to you if it's public or not I'm not sure. You are now the chief architect and minister of lobster affairs. That's a bit of a joke staff for product and engineering. Dave is the chairperson and also you have four other people responsible for community partnerships, talent and finance. You have core maintainers and of course a global community behind you guys. But are these people enough to handle today's open claw daily operation? I think as we scale, we will continue to scale and we'll continue to sort of grow the team and continue to grow like our support system around that. You'll notice like we have a number of job openings currently And we're going through a hiring process to continue to scale the team. But we're also cognizant that we don't want to just explode and hire lots of people because we feel like we need to. I think there's like being mindful around... what that looks like. So for example, if we hire someone into a function, it's like, okay, how does an agent first look like in this role? like design or events. Like how do we build like an agent first version of this and sort of build that up with that, with that individual and then see how far we can scale that. And it's not about not hiring people, but it's about creating efficiency. And then also, giving this back to the community as well. So everything that we do internally and the way we operate, we open source and we give back to the community as well. So I think there's some interesting things like, yeah, we'd love to hire everyone. [55:00] We possibly can straight away, but we also have to be mindful of like cost and like what does that look like and the impact that we need to drive from that as well. So is it hard to hire people nowadays? Because there's apparently a talent war in Silicon Valley. I think for us it's very different. Yes, there's definitely a talent war. But... I think our project is very unique in terms of its position and its mission as well. And I think that attracts a certain type of people into this project, which aligns with our mission and what we're trying to do. So I think it creates like a very unique sort of opportunity. I think people that work in open source, do it for the love I think there's a passion that you want to give back to the community. And I think that we want to attract those sort of talent as well and foster that sort of talent as much as we can. Cool. And you don't always agree with each other, I guess, within the foundation. So when there is a disagreement, who has a fantasy? We've never actually had to put the hammer down and get to a point, but I think eventually we'd work out who said what. But we have enough people and enough differences of opinion that collectively we finalize and agree towards something. And I think a project of this size and complexity, I think disagreement is a good thing. It means we're thinking through the changes, we're thinking through our opinions and our differences. If we all agreed with everyone, I think we would just end up receiving product that just doesn't work well. The fact that we have such a diverse mix of products maintainers on the project and organizations we work with means that we get very challenging opinions. For example, someone wants to put a feature in, maybe for enterprises that helps... in like a, say, US technology space, But then maybe some Chinese maintainers are like, hey, this isn't going to work for us. So how do we then... create a feature that works for everyone or vice versa. It's just an example, right? So it's like making sure that it's fair and works well for everyone is like part of the, part of the challenge. But I think disagreements are healthy. We haven't had any sort of stalemate or any issues with that. I think, open source communities tend to be open source projects, tend to be I don't know. have its own way of solving these problems. Sometimes they can get quite heated, like we haven't had anything too crazy, but they tend to have their own sort of rhythm and flow and way of dealing with these issues. And they have done for quite some time. So a lot of the maintainers on the project have worked in large open source projects as well. So for a lot of us, it's like we're kind of used to dealing with these sort of situations. I have another question I have to ask about Peter. Of course, I know people at the OpenClaw Foundation, also the community were happy for him after he announced that he will join OpenAI. But OpenClaw's early development tied closely to Peter. And now the foundation per se, how can you actually guarantee the mechanism that can work individually. I'd say we don't have like a single person risk at the moment. We have an entire team. of engineers if you look at the commits for example it's like starting to smooth out a little bit more as well you know peter's vision definitely matters a lot to us and the team as well but as we continue to scale the team i think it's how do we take that vision of peter myself and other people like how do we kind of bring that together into a cohesive like strategy roadmap and get everyone aligned to the end state that we're trying to go towards so i think that's the sort of thing that we're working towards that reduces any sort of risk around like it's a one-person team i think a group a lot of the community was still thinking hey it's just Peter right but I mean, fundamentally, to an extent, a large percentage of the commits are still coming from like a handful of people, myself included. But it's hard to quantify because like just looking at commits is not enough because some of the changes can be quite fundamental, let's say security related, but it's only like a small volume that could have large impact. So we have a mixture of people. contributing in different ways. And it's hard to quantify what that looks like. But I would say that it's now quite blended. I might not have the full picture on everything, but not every person on the team has the full 360 degree view. But we're collectively owning parts of the problem and working together to solve it. Yeah. Cool. And looking ahead, what is the single biggest challenge facing the OpenClaw community today? I wouldn't say so much challenge of the community. But I would say... There's a lot of opportunity there. there's a large percentage of users in the world that still haven't fully adopted AI yet. For us, it's like, how do we make this more accessible to the general public, to moms, dads, uncles, aunties, whatever? Like, how do we make AI easy to understand and work with? Because I think if people can access it, they can use it, they can understand it, and it means that they can have a say [60:00] what their future looks like with artificial intelligence. Whereas if you don't get a chance to use it, I think that's not fair. So we're trying to make sure that everyone can get access to something like this technology. And I think there's still a huge percentage of population that still hasn't touched AI. So how do we enable that? On top of that, I think one moment people are waiting for is Apple. People are really hoping that the mobile agent can really work people who are having phones can actually... Most people's computer is their phone, right? Everyone has a phone, but they don't have a computer, so having that on-device agent-first experience is probably going to be quite critical. You guys have any plan to perhaps work with Apple to make the AI experience better? Time will tell. Yeah, but it's hard. We want to integrate with every AI technology out there. But, yeah, we'll see. And open source projects can be more difficult to sustain than the traditional software projects because infrastructure, talent, maintenance, operation all takes money, right? Why would you guys want to keep OpenClaw as open source community? Because we can see... Some open source community, for example, VIM and SG-Lane, the founders actually came out to start their own business, own startup to raise money. They have very impressive fundraising and, you know, they are saying only startups can make things happen faster during the AI era. Do you agree? I mean... Yeah. I don't know if there's a right or wrong answer for this, but I would say Peter's wishes were like, he thinks the only way the open clause survives and is fair to everyone is that it remains in a foundation and it's not controlled by one company or one country or one entity. And that way, it can be truly embraced by the community. And I think by doing this, we've been fully embraced by... China, the US, tech companies, moms, dads, like everyone, right? And I think that's enabled everyone to go, okay, this is like a fair, kind of like a Switzerland of AI. we're happy to work with that. And we've had people reach out to us that exclusively want to, find a way to work with us because we're in a foundation model. So I think it brings benefit. And Yes, it doesn't generate revenue or create sort of capital wealth, but I think it gives back to the community. What we give back is... creates opportunity for the industry to then create more capital. So I think it's a positive sum game for everyone at the end of the day. It's selfless, I think, in a way for Peter to go, hey, I want to give this to the world. But I think it's a beautiful thing that's happened. Great to hear. How is your experience with WAC this year so far? Yeah, I'm just hearing lots of crazy things so far. I haven't had a chance to go out yet this week. But I think it's going to be great. I'm giving a keynote about my journey through open source. which I think is going to be great. And I'm going to be like really excited to see a lot of technology companies here and get part in the conversation as well. I think it's going to be like a very invigorating process week I think by the time the conference is finished I'm going to be like yeah I'm done, but really excited. I'm glad to be here back in Shanghai as well. Okay, cool. Last question. So for the next 12 months, if you can give us some predictions, you know, what are the most important three things that will happen, whether it's agent related or not? I don't want to give too much away. I think the thing that people need to watch out for, I think it's going to be interesting, is like how do agents... Communicate and collaborate with other agents is going to be the next... sort of interesting thing. If, say, a SOTA model company comes up with a solution, another model company comes up with a solution, how do these all coexist and work together? We're going to have a world like phones, right? Everyone's going to have a different phone, but they need to be able to call each other, right? It's going to be a very similar process. So I think that's going to be one that's going to be really interesting to see what happens. We're experimenting on some ideas around that, but that's going to be really cool to see because I think that's going to be where we're going to start seeing My personal agent can call my work agent and share some information and get to work together and solve some problems right yeah i already heard some scary uh stories saying like the ai agents actually invent their own language that humans don't understand We'll see. But a lot of the mechanistic interpretability stuff is quite interesting around evaluation. The agents will find what's mathematically most optimal or, like, whatever's been trained into it. I think we just have to be careful, like... seeing that because agents are doing some behavior that it becomes we don't want to kind of say because x is happening it's y um we need to understand like what's causing it but i think safety is one aspect but also like Just even getting these agents to communicate with each other is going to [65:00] the challenge right like you've got enterprise you've got different software systems you've got different architectures so i think a lot of those kind of challenges are like actually way more complicated than security for example yeah yeah all right okay anything else you want to add no i just want to say thank you for the opportunity to come and have this interview and this conversation so excited to be here Okay, thank you so much Vincent. Hope you enjoy your time here in Shanghai.