OpenClaw Enterprise: Making AI Agents Safe for the Enterprise

0:07Good morning, good evening, WebYar, and welcome back to the Enterprise AI Show. This is your host, Aaron, and today, Brian sits down with two distinguished engineers at Red Hat to dig into OpenClaw Enterprise. They explore what it really takes to run AI agents safely inside an organization, and they explain why the right starting point is to treat every agent like malware. And how a sandbox can actually ask an agent why it tried to do something it wasn't allowed to. From there, we finish out the episode talking about how a project only a few weeks old already has more than 3 ,000 merged pull requests.

0:49And all of that is coming up right after this quick break.

0:57Today's show is sponsored by Nord Security. Your employees are using AI every day. But do you know which tools they're using, what data they're sharing, or where their AI agents are connecting? you can't govern what you can't see. Introducing NordLayer AI Scout, the visibility layer for enterprise AI. AI Scout gives you visibility into AI activity across web and desktop apps, detects sensitive data exposure, helps control AI agent connections, and creates an audit trail for governance so you can protect your data without blindly blocking innovation. See what your AI is doing, then decide what happens next.

1:31Connect with Nord Security to learn more about NordLayer AI Scout. visit nordlayer.com slash ai-scout slash enterpriseai and get started today

1:46and we're back and folks if you were listening a couple weeks ago we were doing a cloud ai news of the week and brandon and i talked about a new announcement that came out from open ai as part of open claw nvidia and red hat called open claw enterprise and i mentioned that i was going to have some folks that we're going to dig into it a couple weeks later from that. We are back to dig into it. And so I'm excited today to have both Derek Carr and Rural Patel from Red Hat joining me. Guys, why don't you guys go ahead and introduce yourselves because you will do a much better job than I will. Derek, I'll let you go first.

2:17Yeah, sure. So thanks for having us, Brian. I'm a senior distinguished engineer here at Red Hat who's been working in the cloud native ecosystem for, I guess I'm getting older, but over the last decade plus. Background traditionally had been And then technologies like Kubernetes and making sure all the world's apps can run on that platform well. And then, of course, agents are the next class of apps. So excited to talk about it with you here today. Go ahead, Bruno. Hey, I'm a distinguished engineer at Red Hat. I've been working alongside Derek on containers and Kubernetes for a long time. And now excited to be enabling Kubernetes and OpenShift to be a platform for AI agents.

2:57Yeah, very cool. Very cool. We're going to do a little bit of history, although it's only about six months or five months worth of history before we dive into this. Back in March, for those who've listened for a little while, we had Sally O'Malley on when a technology called OpenClaw first came out. So it was one of the very first widely deployed open source AI agents. And Sally came on. She was one of the first people who was part of the project in terms of being a contributor and a maintainer. We talked about it at that point. And then there was a big buildup around OpenClaw. A lot of people were talking about bringing it into their workspaces.

3:32And then I don't want to say it got a black eye, but I think the overall concept of, oh my gosh, employees are going to start bringing in AI agents that enterprise IT has no control over is potentially a security risk. Where did you guys start getting involved with starting to look at trying to make AI agents more, I don't want to say more suitable for the enterprise, but more acceptable and more manageable for the enterprise? How did you guys start getting from all the work that you've done around Kubernetes and containers and helping people deploy applications that way to start looking at agents in some sort of context?

4:07Yeah, I'll jump in on that. Probably at some point at the end of last year, Manol and I were looking at where the ecosystem was going. And even prior to OpenClaw, we saw trends changing, right? People were using AI agents to write more of their code. And that code would be oftentimes written dynamically on production systems with a lot of review. And that gave us some anxiety about what do we do as a platform to protect that in the future. And when OpenClaw came out, it basically showed the world what was possible with an AI agent, right? It was an AI agent that could actually do a thing and didn't require a lot of deep technical knowledge to demonstrate powerful capabilities.

4:51So for us, it was like a wake up call, right? To be like, what are we doing to make sure that as these class of applications diffuse inside the enterprise to make sure the platforms we provide are safe? So Rinald and I were talking a lot around what type of evolution to Kubernetes and sandboxing generally is needed to constrain applications like OpenClaw so they can be more least privileged and more safely operated in a company environment. That led us to looking around the community, engaging with others, and then ultimately getting involved very early in the OpenShell project with colleagues at NVIDIA.

5:32And the whole purpose of that project was basically like saying, how do I put an AI agent, whether that was OpenClaw or anything else, in a deny by default sandbox? Totally locked down. You can't do anything unless I tell you you can do that thing. And so that's been a fun activity that he and I have been involved in probably the last six months, right? And so I don't know if you want to share more about your perspective sandbox in there. Yeah, so I think when we look closely at agentic workloads and compare them to containers, the threat model is entirely different. You have something that has intelligence inside through LLMs, and it can do things.

6:12It's not deterministic, right? So you do not know ahead of time what's possible, what it's going to do at runtime. So in the past, we had controls that were deterministic. Okay, I give it these capabilities. I give it these permissions. It runs as this user, and I've tested it. The web application is not going to dynamically change what it's going to do, right? With agents, that changes entirely. So you have to be more paranoid. You cannot give it any credentials. So that's the first thing OpenShell does. So OpenShell injects the credential from outside through a supervisor egress proxy. Your agent running inside the container never sees any credentials directly.

6:58It only has placeholders that are replaced by OpenShell. So that's one idea. The other ideas are how can we use newer kernel features like landlock to further add more security to what containers already do. To restrict what the agents can write and what it can do. For example, it shouldn't be able to write anywhere in the file system. We control that. One more advantage of OpenShell is that it scales up and down. Like you can start with OpenShell on your laptop using Podman or Docker as your sandbox. And then you'll soon realize that, oh, OK, I am not happy running just one agent. I need to be running way more.

7:43And then OpenShell allows you to scale into Kubernetes. So it has compute drivers that allows you to scale into the cloud. I think generally that as a part of our work together, along with our community friends there at NVIDIA, we evolved that project looking at it from treating the agent as almost as if it was malware, right? and having that perspective on it, wanting to understand what the application is attempting to do, and then in some ways make OpenShell a thing that an intelligent application can talk to, to express why it was trying to do that thing. So if you throw a codex or a Claude or an OpenClaw or a Hermes or any of these things inside an OpenShell sandbox, and the sandbox denies its ability to accomplish a task, like it's an interesting novel space to then ask what type of end point should open shell make open to itself that explains i blocked you from doing that but can you tell me why you were trying to do that so then as a human actor i can make an intelligent decision to say i'll allow it down or i'll allow it never and so i think going back to your question then brian was like i think together we felt like we had found the right primitive to make it safer to run this type of application on enterprise compute.

9:04And then the next question was basically saying, how do we make sure harnesses in the broader AI agent application space actually start consuming these primitives? Can we make two communities work well together by being aware of each other? Yeah. So when we first start engaging in a project like open shell, it's like you treat the AI application as a, as an unknown quantity and you just throw it in the sandbox. And basically anything inside there, you can't quite trust. Which meant that there were a lot of interesting features in these AI applications that you are discouraged from using, right?

9:42Memory, I kind of want to trust memory, right? It's holding data. It's an enterprise record. It has to be held under data retention concerns. If I throw the entire application inside the sandbox, basically all elements of it have to be considered untrusted. And so a lot of these AI applications that people get familiar with today that they run on their laptop, Rinald and I started to observe that we think that they were going to evolve. They're going to start to decompose. Parts of the applications will be separated into discrete processes. And some processes will say this is where the untrusted agent loop is versus this other process is where the gateway communication channels are.

10:26So if you speak to your agent over Slack or WhatsApp or Telegram, that's a trusted process. It's only when you gave it to the LLM, the tools, that it needed to be sandboxed. Earlier in the summer, when I started looking around and we were saying, how do we drive adoption of projects like OpenShell? It's no fun to build a tool if no one consumes it. And OpenClaw had been on our radar because obviously it had a lot of popularity, a lot of capability. And so we started to reach out to the maintainers in that community to say, what could we do better together? And one of the things that caught our eye was there was an enterprise RFC put forward by the OpenClaw team that talked about how they wanted to decompose OpenClaw itself and make it more turnkey to onboard repeatable agents.

11:21so you don't have to think about it the same way we don't think about deploying a new application on Kubernetes. Instead of being so scary, there's no good practices, no good processes. And as a part of that, we thought this is very compelling. This is very interesting. We reach out and we say, how do we work together? And like all good open source projects, two communities can start having people lean on each other and say, let's build together. And that's ultimately what led to us starting to engage on OpenClo Enterprise. was. Interesting. So in essence, you guys, if I'm understanding this, you guys were working on, you're working on OpenShell.

11:55You went from how do we, what's the right sandboxing technology? Then it was the thing in the sandbox is untrusted. And then you went through the exercise of being like, that's not exactly how agents are going to work in reality. Like you said, there's some things that are sort of inherently trust. We try to actually use the agents and then we see, man, there's certain capabilities I'm having to not trust that I want to trust. How do I go and get agents to build differently. And then OpenClaw is sitting there thinking from the other perspective. They're coming at it from the agent perspective saying the same sort of thing.

12:23If we want our stuff to be widely deployed or we're seeing these use cases, we're going to have to figure out how to be more modular. And the two of those sort of concepts came together this past summer. Is that? Yeah, perfect. And then the other, obviously our background being in Kubernetes and stuff was there was a very clear deployment target of saying we want to put together the best practice for deploying claw-like agents at scale on Kubernetes. And that's our background. That's our sweet spot as well. And we thought for the last 10 years, we've been trying to make sure all the world's applications can run well on cloud-native infrastructure like Kube, whether that was distributed web apps, stateful web apps, packet processing, to now with a new class of agents that have emerged, to some degree, it's no coincidence that these things need to co-evolve together and you actively have to reach out and find those on the frontier, kind of showing the world what agents can do to then figure out what the problems are and then how we work together to make them run well on the infrastructure we have.

13:27Unlike in the past, everyone knew what a distributed web app largely was and we saw an unknown good pattern and we then try to make sure Kubernetes can fit it. Here, we need to co-evolve together, which also just means we got to get dirty and start working in those communities together. And that makes sense. Obviously, what we saw in the past with you had Docker as sort of the container technology, kind of working independently, you had Kubernetes going, okay, we want to do this at scale, those could be somewhat independent, obviously, there became a lot of connective tissue between container runtime.

14:01But yeah, as we get into applicants, as we get into agents, which be some become sort of applications, there's probably a different level of sort of granularity that you want to provide from sort of day one, as opposed to just being like, any and any, and they'll figure it out and go from there. So I'm curious, the three of us spoke last week, we were doing a little bit of prep for this. And before you guys were telling me what was happening, I had this mental list of things that frustrate me about agents in the enterprise, as opposed to just agents that I run on my desktop or that I use for regular use or whatever.

14:34And my list was pretty short. It was pretty simple. It was, where can I run an AI agent? Because maybe I have the right hardware. Maybe I don't have the right hardware. We've seen people buying Mac minis and all sorts of stuff because they want to isolate stuff. The second was, what's the starting point? There's lots of jobs within any organization. There should be some starting point. So you're not just being like, here's a big blob agent. You can do anything with it. Setting up MCPs, getting into tool calling and external stuff. I felt like it was difficult to share skills or experience. I've got plenty of coworkers who are like, I've been working on this thing for months and I'm really good at it.

15:08And you were like, how do I get up to your speed quickly? So difficulty in sharing stuff, again, everything was at the individual level. There really wasn't a team concept around AI agents. It was somebody might own a swarm to go do a task, but there really wasn't a team thing. And then the sort of going back to the original thing of how do we separate personal from work, which gets into sort of isolation stuff. And as you guys were beginning to talk about it, I was like, okay, this stuff is going to check a lot of those boxes. So So why don't you start walking us through what is OpenClaw Enterprise?

15:40And maybe most importantly, what are some of the most important things that the overall system is going to start bringing to the market that people are going to be like, oh, okay, cool, this is going to really move us forward? Want to jump in on that one more? Yeah, sure. So I think first you said how people have their own agents. Within many enterprises, you'll see folks have come up with their own bespoke agents, but we are not speaking the same language. Right. Like when we pick OpenClaw as your agent solution, everyone now has the same language. They know what they mean when they say memory.

16:15They know what they mean when they say soul, identity, skills, MCP server, plugins. So now we are converging on what the standard definition of an agent is with OpenClaw. And now we want to scale it out. So with OCE, the pattern is actually very similar. if you're familiar with Hypershift. So Hypershift has hosted control planes, right? So OpenClaw has something called as a gateway. So with OCE, you can imagine there are two clusters. So there is one cluster, which is a control plane cluster that hosts your gateway. And then there is another cluster that hosts your agent harnesses running within sandboxes.

17:01And the OCC control plane is basically a way to get your own agent. And under the hood, it's setting up a gateway for you on the control plane. And that gateway is communicating with an agent that runs within OpenShell on your data plane cluster. We talked about how, like when you put everything in a sandbox, where does your memory go? With this architecture, we can start treating your sandboxes as replaceable. If for some reason the sandbox dies, all the memory, all the session database information, all of that is stored on the control plane side. So the gateway can just go, oh, I lost that. Harris, no issues.

17:46I know where it had reached in its turn. I'm just going to spin another one, give it all the context, and it'll pick it up from there. So that's the basic idea. And I think onboarding generally, OpenClaw Enterprise includes a web console, an API, its own control service, a store. And similar to Kubernetes, where people create namespaces and then render pods and deployments into there, OpenClaw Enterprise has an equivalent project concept, which it ironically does call namespaces. And you create one, you go inside that namespace, you say, I want to create an agent. One of the nice things about the project's goals are, you touched earlier, that some people can become what feels like wizards with these tools, based on what they can craft locally on their setup.

18:32And it's hard to share that knowledge with others. Versus you'll come into OpenCloud Enterprise, you'll go to hit Create Agent, and you'll see a list of known good presets for agents of particular business goals or outcomes you want to achieve. So if you wanted an agent that could act as like a question and answer assistant for your team, you'll have a non-good preset. If you have an agent that wants to help you do team software engineering, co-development, you'll see a preset with that template. If you need a simple personal assistant or chief of staff to support your work, you'll see an agent with that type of preset.

19:08And then the benefit of that is anybody in your organization, as they configure their agent for their use case, can then save that as a repeatable template for others to follow. So we would expect with a project like Open Enterprise that agent templates for particular business functions, outcomes, or roles will just become standardized as, oh, I'm using this known good configuration that's been blessed by my organization to that known good task. From a dangerous standpoint, one of the benefits of OpenCloud Enterprise is everything about it is policy-confined. So by default, when OpenCloud had come out, there wasn't necessarily as many security controls as you'd like, and it was unknown in advance what that agent could go do.

19:53When you deploy OpenCloud Enterprise, one of the goals there is you'll have a ceiling that defines the maximum allowed permissions or network endpoints or programs that your agents can use. So that you as a company who wants to put this into their organization, you can set that permission ceiling and you can feel safe and comfortable that none of your employees will go beyond that known good action set. But then when your employee population comes in and says, I want to create an agent to help support me in my job, that agent starts totally least privileged. And then based on the need of the job, it will be able to grow privileges up to that ceiling, but not beyond that.

20:37And so that's really extremely important, which is if you have these applications that take unknown input, right? Like random user prompts, random questions, that type of thing. You want to, you can't really audit them in advance to know all possible things that would happen. So if you know there's a maximum operational surface that they can fit, you can then box your employees and their assistants in that known good set and then make it the outlying behavior when they say, oh, I'm trying to achieve a thing. Can I get permission for that? Like at that point, you can then review it, ask if it's safe or not safe and that type of thing.

21:13Yeah. And then on working together, Mernal and I work on a lot of open source projects together. And oftentimes with AI tooling these days, we'll find ourselves in a room like this. And we're just looking at a shared coding session, right? And just one of us has the keyboard. And it's like pair programming, but the other person's just kind of chiming in, but not getting to guide the keyboard. One of the really cool capabilities that's in OpenClaw now is you can set up a claw to be multi-user. So for example, if you wanted to set up, and we've done this now in the maintainer community here, you can set up an instance of your agent that allows your trusted co-maintainers to collaborate on an outcome together.

21:59So if the three of us are in the instance, I can see Brian, you're in there chatting to the outcome that Rinald and I are working on. And we could be working on, let's say, a pull request somewhere or a code change or a document of some kind. And each of us can be feeling like we're a part of the authorship because we all get to be at the keyboard. And rather than just seeing the output thrown over the wall, each of us knows the collective input that led to that output. I think personally, we think that's going to be extremely empowering for organizations that are trying to figure out how employees work together with AI rather than work in isolation with their AI.

22:39Yeah, I think in that spirit, when you talked about you learned a thing, but it was hard to share that with others. It'll be way more visible to everybody like Brian, why you are really great at AI versus someone else when everyone can see how you use it right together. Right. Yeah. And I think that path, there's always going to be, especially in a work environment, there's always going to be a certain amount of tension between I want to get my work done. I want to do it the best I can. And maybe you get rewarded accordingly. But at the same time, most work within some sort of company tends to be group work.

23:11It tends to be teamwork. It tends to be new people are added, people leave the project or something like that. And so if you don't have, whether it's the equivalent of memory or whether it's a shared mechanism to see how you got from A to B, it does make it, it does put a cap on how productive groups can be with AI. So I'm glad that you guys are starting to see, starting to build frameworks around that that are going to be fairly easy to access and kind of work within that start to make that piece of it easier. If you think about it, like open source communities have traditionally just been one big group project.

23:43Yeah, yeah. And we all would toil together traditionally in isolation to create some code asset, throw it over the wall, look at a pull request and say, give me your thoughts. And all you could see was the result, but you didn't necessarily understand the motivation. Part of the appeal to, I think, Rinaldo, and I'm curious your perspective on this is like, as technologists, like we see how we work is changing. And like, honestly, like working in a community like this to see how it can still support open source and open source maintainers working well together. It's really neat to just be on the forefront of a tooling ecosystem that's exploring how do we do the world's biggest group project, right?

24:24Because if we're successful there, every other project of a smaller scope, smaller user population can probably learn from those patterns and we can get the right right UX in place but I don't know how you feel about that but yeah absolutely for me I feel like with the advent of AI you've seen a lot of open source projects struggling with the influx of PRs because it's become extremely easy to put PRs but you don't necessarily understand the motivations and the end goals that you're trying to achieve right what people are doing is they are chatting with their own agents they are creating prs and maintainers are not able to keep up with this new development model like derek and i can work together we can chat to the same agent so we are not like slinging back things back and forth with our own agents and instead of having multiple agents we are talking to the same entity that has the context from where each of us is coming from and once we reach agreement it becomes way easier to drive that single agent with the entire context to completion, right?

25:30Versus the fracturing and the back and forth. So I really feel that's the future of development, both for open source projects and also for proprietary stuff. Yeah, no, I think it's good. As I'm thinking about this, I'm trying to think about both sides of the equation as we start to talk about like enterprise stuff where it's what's happening on the operation side. So it seems like this is going to be technology that's going to natively integrate with Kubernetes, which lots of companies have in place today. They already run lots of applications there. So you've got something that's going to natively integrate with stuff that you're very likely using.

26:06You're going to have, as the operations team, you're going to have the ability to control or manage the control plane. So you're going to be able to help with scale. You're going to be able to think about security at a higher level. And very likely, you're going to be at least part of the interaction with some of these default templates, like you said, where you might do a double click on something and all of a sudden there's 100 different jobs or 100 different skills, right? So we're giving the IT organization, the operations organization, the lines of business, levels of control that they need and levels of security they need.

26:39And on the end user side of things, which is always where we have friction of, I want to be able to do anything I want to, IT kind of shuts me down. It sounds like the work that started with OpenShell is starting to evolve here with whether it's collaboration, team stuff and sharing. you're going to be able to you might end up starting with a starting point agent that that is somewhat shut down because by default things are kind of turned off but as you start to work there's going to be a a mechanism and i don't know exactly what that'll look like that allows you to be like hey i tried to do this thing it got shut down i assume you'll then be able to pass that along to somebody and be like hey i'm trying to do this legitimate work or this is why this is why it started to do those things and it feels like there's going to be a little more kind of collaboration or two-way channels between the people that want to get things done with agents.

27:28They're going to have to do them in a secure way and being able to integrate or talk to the operations teams and be like, this is why I need to do this stuff. This is why this stuff is legitimate. Can you open up these access to tools or open up these other types of things? And it feels like maybe there's going to be a better sort of collaboration than we've had in the past between end users and IT groups who always fought about, this is the strict policy, I can't break it. Yeah, I think anybody who's used tools like Codex or Cloud on their laptop is familiar with the question of, do you want to give a permission to do this action?

27:59And you can very quickly hit a yes loop, right? You're just hitting enter over and over again. To some degree, there's a lot of trust then put on every user to basically be on their game and make sure that what they're accepting is good to get accepted. But we all understand people get tired, right? And they can sometimes make choices out of convenience or confusion. That's not always right. But because many of these things are running on their individual devices, it's easier to look away a little bit and hope for the best. Once you start running these things in your data center and they're on your network and your VPC and have the ability to potentially touch other applications running co-located to them, the game gets much greater.

28:42The nice thing about OpenCloud Enterprise and the repeatable patterns there is I don't think we're known. I think every possible agent in the world will be fulfilled by this one particular project. Let's say there's not all agents need to be persistent and or long running. But for those that need those agents that you want to be able to talk to over Slack, those agents that you want to be able to refine themselves and get better over time and learn with your usage. that class of application we think is going to be well served by this control plane that's being built here so that the ceremony your operations team needs to do to onboard a new one is trivial and well understood right and we want to make this part of the application process that had been dangerous more routine so that then the outlying scenarios where you have a more bespoke agent for a more bespoke use case can get the scrutiny they deserve.

29:40Like right now, everything's thrown into one big bucket and figuring out how to security review or analyze an agent that you want to onboard in production is actually a very daunting conversation for CISOs and teams to have. We think we'll be successful at carving off a large swath of that with this effort. Yeah. And I have to imagine as far as, like you said, long running agents or long running processes and so forth. That's going to apply, especially if it's things that you're trying to deploy and then have it get better over time. That becomes everybody's personal agent because you want it to know about your work evolves and so forth.

30:17But it's also going to be just a ton of things that might not be tied to an individual, but are going to be tied to things that improve how you interact with your customers or improve individual groups that just do repetitive processes and want to get evolved over time. So, yeah, I can imagine, like you said, this is going to hit a really broad swath of enterprise assistant use cases, workflow use cases, and other types of things as well. On that point, I'm not as young as I used to be, right? None of us are. And both Bernal and I have worked in the container and Kubernetes space for well over a decade.

30:50And you accumulate a lot of knowledge and context in your long-term memory on that. And our colleagues will ask us all sorts of questions. How does this work? Why was this choice made? Why is it like this and not like this? And if I think back how things would look differently if the technology that we're talking about today existed 10 years ago, and it could have been present for every step of the journey, my memory is sometimes poor. And I could forget why this knob was set to X versus Y. But the moment you have a team agent that can act as your shadow to help you recall that which you forget, I think things get much more better and much more interesting.

31:36And one of the things that we're doing in the maintainer channel for OpenClaw Enterprises, we just rolled out our support agent for the project there. We call it Skipper. And we'll say, oh, here's a change. and then it's not infrequent that we're just adding Skipper to say, go and ask a thing about it, tell me how this worked, or is this in conflict maybe with something we discussed earlier? And it just helps you focus on the problem at present without feeling the burden of having to remember everything poorly that happened in the past. And I think in a team setting, that's going to be very interesting.

32:12We deal with teams that need to onboard new people when people go and take new opportunities. Right. And so imagine you're a support organization and it's your first day at the job. Typically, you go and find the most senior person and you're asking them a ton of questions and you feel like a pain. Right. Now you could have an assistant to go and help answer that to get you up to speed super quick. That's so empowering for everybody. I think that type of thing will be more diffuse broadly going forward. Yeah, behind every code artifact, there's like tons of Slack discussions and context that get lost.

32:49And now, hopefully, if we get the memory system right, that context will stay on. Like, even if we move on to other projects, someone can come and ask, oh, why does OCE work this way? Hopefully, the agent will be able to answer, okay, Derek and Mural had this conversation back in 2026. And this is why it works that way. Yeah. Yeah. There was one more thing I know we talked about a little bit before, and I know I talked to, as I talked to a lot of different companies about it. Right now, a lot of the generative AI engagement companies have is with something in the public cloud. Maybe it's they're getting it from AWS or Azure, Google, or it's a specific agent from a specific harness.

33:32And part of that is that that was something they picked and they just went to. But some of it is we didn't necessarily have the money to buy GPUs or we didn't want to run GPUs. This is a little different than we had in the early private cloud days. And then people picked where things worked best. Did they run best in the private cloud or the public cloud? how does this stuff like how tied to this I know we talked about this being open source and there's community with it but how tied to this is it to is it tied to a specific cloud is this something people can run independently anywhere they want to do they have to have their own GPUs to make this work or what's the where's the what do I have to have versus where can this run anywhere or be utilized anywhere I'll start on this and then Manoli can extend it's very important to us that the technology we work on is open source as in we all traditionally think about it.

34:20I want to see the code. I want to be able to contribute the code. I want to be able to participate in its co-evolution. So the first project we were working on here, we discussed around sandboxing generally, that open-shell technology is open and just was recently donated to the Linux Foundation as a part of the CNCF. So the thing that would be securing your AI agent in its untrusted actions, you can inspect, you can evolve, and you can help us co-maintain together, right? Then if we go up a layer of the stack here with OpenClaw Enterprise, one of the things that is appealing to us is it's backed by the OpenClaw Foundation itself.

34:55It has a commitment to maintain its open source posture and its overall neutrality in the ecosystem. So you can run this anywhere. You can run it on your laptop. You can run it on your data center. You can run it in your cloud of choice. that you get to choose as the builder, how it's presented, how it's configured and how it's tailored for your use case. So that control element is entirely there for GPU enablement. Yeah. If you have the GPUs available and you want to run a model of your choice, whether that's an open weight model of a large size, small size, what you need for your use case, you can entirely configure the application to use that model.

35:38There's nothing tied to a particular their frontier provider. And in fact, there's an extensive amount of custom visibility. You can say which model is used for which element of the task. There's interesting features in the ecosystem around memory and dreaming where you might choose to bias for one model or the other. So yeah, one of the, obviously the attractive appeals for us here is that, yeah, this is technology that is open and visible to the world and you can run anywhere you see fit, connecting it to any infrastructure fit as well. I don't know if, Maui, you want to add anything else? Yeah, yeah, exactly.

36:13So if someone is using Dynamo or LLMD to host their own models, you can totally connect the system to user inference from there. You can keep it all within your own compute. You don't know how to use anything on the cloud. Yeah, and I think that becomes, again, we've seen lots of technology trends. We've seen there's always an early adopter that tends to be a single vendor, a single stack. a bunch of vertical integrations. And then over time, we tend to see open, whether it's community based or open source or multiple choice tend to be become more widely adopted. And I think, especially given where we are, we have this intersection of new technology, but we also have a lot of geopolitical things going on.

36:53So you've got concerns about sovereignty in Europe and Asia. You've got people working with different vendors, depending on what's going on at any given time in AI. So I think that piece is really important. I mean, it doesn't get talked about enough. But yeah, having the ability to control this and be like, hey, if I want to keep dealing with it in the public cloud, cool. If I want to do it on private or some combination, that's available too. Real quick, I know this stuff has been fairly recently launched. Obviously, we talked a little bit about this is building upon some other projects that are maturing, whether it's OpenClaw or OpenShell or some other stuff.

37:26Do you guys feel like as the community is growing, we talked about Red Hat's involved, NVIDIA's involved, OpenAI's involved. Is this something you think we're going to see quite a bit of rapid maturity as we get into, obviously, the end of this year and next year? Is this something you expect people will start deploying probably into production in 2027? Does it feel like it'll be at that level of maturity? Yeah, which is insane when you think about it, right? We're talking about two precursor technology projects, OpenClaw itself and then OpenShell that are both less than a year old. Right. And the, as some of the earliest contributors to Kubernetes here, I remember it took about 13 months to get to a cube 1.0 date.

38:08And the, the ability of that project was much different than what we appreciate now, right? It was a far more constrained MVP. So for OpenClaw Enterprise in particular, it's been driving a lot of nice features to OpenShell that will continue to round out the use case so we know what's missing on the sandboxing side. For OpenClaw Enterprise, the goal is to get to a 1.0 release at the end of this year, very early next year. And we'll just see how things shake out on that as we get into the holiday season. And then one of the important things in my mind is how much time we have to dog food the platform ourselves to figure out all the edge cases so the experience is pleasant for everybody else who's first touching it, right?

38:54Like, I think we all want to make sure that you have known good golden path journeys to follow through as you first engage with the project. And so that'll be an important thing for us to reason through. But going forward into 2027, yeah, I think I would anticipate organizations will be piloting this and growing it at scale. and rolling in the production for sure. Yeah. And like you said, it's been building on some projects that are relatively new, call it a year in AI timelines. That feels like a long time. And then obviously it's built upon Linux and container and Kubernetes technologies that's been around for a decade plus and in some cases more than a decade.

39:30Yeah, I think it's really encouraging to see open, not only like open source, but large vendors plus open source, which has been a formula that's worked very well in the past. I think we're seeing corporations trying to figure out how to bring productivity to AI, agents are obviously going to be a part of that. So giving them a way to do that is controllable, secure by default, and then is going to give teams more ability to collaborate, bring people online. All of those long running use cases, I think it's going to be really powerful. Guys, folks, we're going to put some things in the show notes.

40:02They may end up evolving over time because I know Derek and Meral and the rest of the community are building some things, but we'll try and get over the next few weeks and next couple of months, we'll get some links in there as far as some good videos to go take a look at, demos and things. Obviously, you can go take a look at what's out in the community today. There's code out there and discussions out there as well. So if anybody's interested in that stuff, definitely go take a look there. Guys, any last things that we didn't cover that we should make people aware of or just give them a good starting point for going to explore OpenClaw Enterprise?

40:31Yeah, I would just say anybody who's interested in the topic, we would encourage you to reach out to the project and contribute your ideas. And there's public Slack forums that are available if you wanted to connect and give back your use case. And what's exciting about this project is how rapidly it is coming into shape. So I think we announced the project maybe the first week of September. It's been a bit of a blur myself. And the project, I think, has greater than 3 ,000 pull requests merged at this point. And on any given day, 100 new things can get in. And it just is evolving rather quickly.

41:07By the time this is even published, it might be 20 % more feature complete than it was just now. It's an exciting and fast-moving project. No, that's cool. Munral, anything else we left out? I think this is extremely exciting for running AI agents in the enterprise. So if you want to come help shape it, join us, work with us. Yeah, that's great. I think for anybody who's interested in being a builder in this stuff, if you ever, there's times in your career when you think, Like, oh, I missed out on something, some big thing that happened. If you missed out on containers and Kubernetes and you want to be part of the next sort of big thing, obviously AI and agents are probably going to be at the core of that.

41:43This might be a great place for you to jump in, especially if you've got background in this or you're interested in helping open source communities. All right, guys, thank you very much. We've covered a lot of stuff, obviously, for a lot of people. This might be new, but thank you so much for the time today. We'll try and do some follow-ups on this maybe every six months or so as the project evolves. I think it's going to be pretty core to a lot of enterprises and large organizations. So, guys, with that, folks, we're going to wrap it up. I want to thank both Derek and Raul for their time today.

42:08And with that, we will call it a day and we'll talk to you next week. Thanks for listening. Check us out at theenterpriseaishow.com for past shows, newsletters, and all things Enterprise AI.