KubeCon + CloudNativeCon NA 2025 | Idit Levine, Solo.io

Clips
More from KubeCon + CloudNativeCon NA 2025

Idit Levine

Founder & CEO

solo.io

play_circle_outline Overview of SoloIO's progress and focus on networking within the cloud-native ecosystem.

play_circle_outline Empowering Platform Engineers: Overcoming Challenges in Scaling AI Operations for Python Engineers

play_circle_outline Introduction of K-Agent to streamline agent deployment in Kubernetes environments.

play_circle_outline Discussion on Agentgateway as a security proxy for managing agent interactions.

Info
Transcript

Idit Levine, Solo.io

Idit Levine

Founder & CEO solo.io

In this KubeCon + CloudNativeCon North America 2025 interview from Atlanta, Solo.io Chief Executive Officer Idit Levine joins theCUBE’s Rob Strechay to walk through Solo’s evolution from Envoy-powered gateways and ambient service mesh to AI-first networking. Levine explains why platform engineering teams are now stepping in to operationalize AI, detailing how the K-Agent project turns Kubernetes into a dedicated runtime for AI agents with the scaffolding, persistence, observability and abstractions needed so Python-focused AI teams don’t have to think in CRDs... Read more

explore Keep Exploring

What is the history and focus of the company Solo in relation to networking and AI? add

What was the call to action mentioned in the context of Python engineers and the development of AI projects? add

What is K-Agent and how does it facilitate the deployment process for AI teams using Kubernetes? add

What are the responsibilities of the three products mentioned? add

bolt Powered by CUBE AI

Idit Levine, Solo.io

search

Rob Strechay

>> Hello, and welcome back to KubeCon CloudNativeCon, North America 2025 from moderately warm Atlanta. We're finally starting to get things heated up in here on the final day. We're down to the last couple of segments that we're going to be doing, and I'm really excited to have Idit, who's the CEO of SoloIO on with us. Welcome on board.

Idit Levine

>> All right. Thank you so much for having me.

Rob Strechay

>> So, kind of give everybody kind of an update on where Solo's at, and things of that nature, and then we'll get into some other stuff here.

Idit Levine

>> Yeah, no, so I mean, we doing so well. It's so exciting. So, Solo is a company, I don't know, maybe we are alive like seven years. We started to solve the problem that we thought will be the biggest problem that exists in this ecosystem, which is the state-led cloud native community, which is basically the networking. So, we started with Gateway, and we basically have a Gateway built on top of Envoy, and then we basically also are the leaders for STO is basically the service mesh. We were the founder together with Google of a ambient mesh, which is basically the site colorless, and a mesh, and that is way more performance, and efficient. And now we as everybody I guess, but a little bit ahead of everybody, we are basically doing a lot of stuff in AI specifically because AI is also a big challenge in AI, of course we'll be networking. So, we had quite a lot of knowledge there to help.

Rob Strechay

>> I mean back when we were talking in London was MCP was all the rage, and things like that. Now you're talking about some agent registry, and agent skills. Why don't you kind of help us understand, because you were on the main stage talking about some of this.

Idit Levine

>> Right, right. So, actually, actually, look, we can go all the way, kind of like when we talked I think in London, I think that we announce MCP, and we needed to educate people what it is. And right now I think everybody knows what it is. This is the biggest thing that exists. I can explain you a little bit the way I see the market, and what's going on, but in the nutshells, we have a lot of customers. Our customers are enterprise customers that are focusing on platform engineering, basically team, they're running that Kubernetes, they're running all those cloud native environment. Usually, or before that, if you look a year ago, they weren't that much involved in the AI at all. Maybe a little bit on inference, but honestly not much more than that. When we talked, I think maybe as you said in London, we kind of explained how we can actually play a role there. Us like the platform engineering, it wasn't really like people, it wasn't in their charter yet, I guess. Right now what we see is that, or at least what we see that's happening in environment right now is that in the beginning all those organizations had a team, the AI team, or the AI analyst team, they got all the money. They basically needed to build those agents to build those to basically figure out everything AI as well as to execute on this. And those people basically went, and they needed to run it in production. There were Python engineers, they never ran anything in production. And then you see a lot of report of tons of money, and billions of dollars is basically investing AI, and no one going to production, and people talking about bubbles, I mean maybe it's a bubble because it's not really useful. I think that, what was my call to action back then in London, it's kind of like happening right now, and I see it's starting happening is basically, guys, we need to own that piece. I mean the Python engineers who's amazing, and writing this amazing AI stuff that I probably cannot write, do not know how to bring stuff to production, they don't know how to do security. All this stuff that we already built it as a community. So, my call to action said we need to own that piece, and this is why we announced basically the MCP Gateway. What we see right now is, and by the way, we also announced a project called K-Agent because forget about one second MCP, or in generally when you're writing agent, you want to run it, you need an infrastructure. We believe Kubernetes should be that infrastructure, and if it does, there is a lot of gap to close. And that's basically what we decided to do as a company. So, we announced K-Agent, which is basically taking this Kubernetes, and make it a runtime that you can run actually those agent, no matter which framework you're bringing, because you want.

Rob Strechay

>> So, it's helping abstract it from the underlying...

Idit Levine

>> You build it, the AI team will build it, which is they should build it, this is right, but then envision the CF push experience. They can basically build it locally, test it locally, MCP server locally, everything is working. It's fantastic. Now they basically wanted to run it in production. So, what we did is we basically created all this scaffolding. It's automatically creating your CRD, it's automatically creating everything they don't know even right for them, they're just doing right now, deploy, and where do they want to deploy? And it's basically going there. So, with the CI-CD process, and so on, but in the nutshells, yes, so just let's understand what is missing in Kubernetes, right? Stuff like persistent for instance, right? I mean when we talking about Kubernetes, we always talking about stateless application, but agent are not stateless, right? You have short-term memory, long-term memory. You don't want it to start the conversation all the time from the beginning. So, the kind of stuff middle, the human in the loop, and so on. So, we basically had all of this stuff that we needed observability to Kubernetes in order to basically make it that runtime. And we also, as you said, abstract that, because those people probably when they're talking about agent, and they don't want to run agent, they don't want think about CRDs, they don't want to think about Kubernetes Gateway API. They don't want to think about Port, they want to think about an agent, they want to think about dual, and they want to think about policy. So, we abstract that as well. So, that's basically what K-Agent is, and we basically donated to here to basically build it together in the community. So, that's basically the first piece that we needed to attack.

Rob Strechay

>> Yeah, no, I mean I think again, you guys have always been, I mean you did that with Ambient back in Salt Lake, and with you, and Google, and everything. And I think when you start to look at how the community has come together, I think to your point right now I did the infrastructure as code conference on Monday night. I ran a panel there, and then I was talking to a lot of the platform engineer folks who were in the room. There was about almost 150 platform engineers in the room. So, having these discussions, and going through a big piece of it was, hey, it's not only difficult but we're missing out on the governance aspect of it as well. How is the agent registry really playing into operationalizing AI, and really that governance piece as well?

Idit Levine

>> Yeah, so I mean, this is what we see when we open source K-Agent, and we also have K-Agent enterprise. So, we're working with a lot of organizations, same thing. We have Agentgateway, and we have Agentgateway enterprise. So, we are working with every organization that you can imagine almost. And talking to them, it was very clear-cut that everybody is going want to, and they ask us if we have it, build a registry. And why is that? Because that's where you're starting. They see all this people in the organization basically going, and trying to use AI, which is makes sense. And if they're not going to give the government for this, as you say, just going to go wild, people will find a way to do this. Maybe not a good way, but they will find a way to do this. So, one of the things that they care of is first of all, let's put this thing in place. Let's first of all have a registry that is basically blessing with the organization that will own all the models that they can use, all the LLM models, every agent that they're okay of using, and every MCP. And now also skill, we'll talk about it later. And the beautiful of it is that now people can share. So, basically I can build one, I can share the organization is basically saying that it's okay, and then other people can use that. So, that's basically what we did it. But again, you need to think about how to do this right? Because there is an official MCP registry, and there's a lot of other registry out there. So, again, by building it, we put a lot of these things in place. For instance, we importing all those data from others, we enriching that with a lot of data, and metadata, and then basically give the repository operator, and ability to decide what they are, basically curate what they're okay with, and why not. And then it can published internally in the private. So, we see everybody seriously, if you would come to a booth, you will see the amount of people who basically was interesting, and it is amazing.

Rob Strechay

>> Yeah, no, I think so because I think a lot of the toil comes from understanding what's available, and you don't want to reinvent the wheel. You don't want five agents that do exactly the same thing, or have the same set of skills to the next type of topic. But when you start to look at it, and you look at Agentgateway, and the agent's skills, how does it really help be more, I guess you could say secure, and deterministic as well?

Idit Levine

>> So, I mean I think as I said, we have three products. Let's figure out what is each responsibility. The first one is it's a run time. You can run agent, it's like Kubernetes, right? It's on Kubernetes, but it's basically act on the same platform level. The end there is the Agentgateway. And Agentgateway is a proxy. And we can explain about why we build it because we used Envoy before. Why did we decided to choose a new proxy? A very good question that we have a very good answer for that. But it's in charge for security, each job, every proxy, that's the enforcement point. Who can allow to talk to you what they allowed to do, exchange what is identity of agent, and how it's play a game. That's everything, and the responsibility of the gateway. And then besides that, there is the registry. The registry, as you said, it's more government, but we inspired a lot by Docker. So, Docker is the registry, and they also have the runtime, right? So, we also build the runtime to this. So, basically it's very easy to start with this to build Agent, to build MCP tools, to push it to the registry, and then basically to pull it from the registry, and run it honestly anywhere. It doesn't have to be in K-Agent because I actually think that people are not going to run only on Kubernetes. I think also in Kubernetes probably on-frame, but in the cloud is Vertiq. There is agent core, there is a lot of other. We believe that people will run everywhere we need ....

Rob Strechay

>> and everything else.

Idit Levine

>> Exactly. We want to present exactly. So, I think the registry's kind of like this point that we'll be able to bring all of this together, all those platforms that you'll consume. And you should, you shouldn't bet in my opinion, on one platform because innovation right now, it's everywhere. You don't want to decided that you're only going on Amazon for instance, and then you're missing everything Google, and OpenAI do, right? So, you do want to make sure that using everything, and I think that the registry is exactly that piece that will be able to help you to bring whatever you want. And still an Agentgateway will make sure that no matter what you're choosing, it'll be secure, and observe, and connect, right?

Rob Strechay

>> And we were even talking to Google about that yesterday with that piece that they're bringing the AI agent sandbox that they're bringing to the open source as well. When you look at it, though, I mean you briefly touched on it, kind of touch a little bit deeper on the agent skills because to me this is that third leg of the stool that really you need to not reinvent the wheel with.

Idit Levine

>> Exactly. So, it's really, it's interesting because MCP is a big word. I'm the one who was pushing it last but it's interesting, because if you're looking at the evolution of tools, let's call it, we started by basically putting it inside the agent code. There was a local tools agent was using tool, and I needed a code when I build my agent to basically say which tool to use he can use. And then basically the agent will decide that if he's using him, or not. But there was a first problem that everybody can think is that now every time I wanted to basically use the same tool, I need to copy paste it to a different agent, or if I want to share it in the organization, even I can't, it's part of the code. How am I going to do that? So, that's where MCP came, and MCP basically said, "Listen first, let's standardize that because everybody should talk the same protocol." If there will be a protocol, it'll be easy. We're all going to be able to collaborate. So, they basically put the MCP protocol, and then they basically also make it charitable because now it's host, and everybody potentially can use that. So, that was great improvement. You don't need to copy paste, but there's still also some things that there is downfall from MCP, or agent, or local agent, and that's the efficiency. So, when we're talking to agent, the most important thing by far is the context window. That's what you're paying money for. So, you don't want it to waste it. So, when we're talking about the context window, when I'm sending something to the LLM, I'm sending the information that he will know how to execute. Today when we're sending in the tool, we are not only sending what called the front matter, which is basically the name let's say, but we are also sending description, and parameters, and return value, and how to call it. And a lot of data, honestly, we don't even know that it will need to call it. So, why are we sending all that data? So, I think that's one of the first thing that basically, and all of this is basically filling your context window. You're paying for this, he's putting a lot of tools besides confusing the agent. You're also basically paying a lot of money for this. So, that's the first thing. And the second thing is there's also, it's called intermediate context that creating a problem, because if you bring in some data, if you have a workflow that you want to do on every step on the workflow, you're going back to the LLM with all the material. It could be files, and whatever. Again, the context with the staff. So, that's what skills basically come to solve. They basically said, can we call it more efficiency? And that's the first thing that they did. So, now that evolution skill is interesting. It's deterministic because it's a workflow, and the agent is basically executing them. So, again, way more efficient, less money to pay, still shareable, because it's basically a file on your operating system. But the beautiful of it, as I said, is that now you are more efficient in the way you're doing this. And the coolest thing is that now the agent can actually call it itself, or write it itself. And that's to me is the blown-mining thing, because agent will be able to actually write his own skill, and execute on them. And that's where the sandbox coming, if that's what happening, and the agent is the one that actually writing those codes, and can execute it automatically, it's amazing user experience for me. But that means that we need to protect it. That's what the sandbox that we use Anthropic, because Google didn't push the code yet, but Google working very hard to make sure that when that agent executed, it cannot harm us.

Rob Strechay

>> Right. So, this has been great. As look out to Amsterdam, and then Salt Lake City again after that, what do you think we're going to be talking about over the next year here?

Idit Levine

>> Oh, that's a very good question. I think that as people kind of delivering, I said at the beginning it was all on the AI team, and now they're starting to kind of give pieces to us. It started with MCP, that's why the registry, and the agents start. But I think that when people will start going to production, efficiency is going to be a big thing. This is why skill is so important, for instance. Sandbox, because security right now, again, those guys are rarely deterministic. We don't know what they're doing. You need to make sure. So, I feel that when we actually start to execute, which I think right now people way more... There's not a lot of production use cases. And therefore when we will start running a production, suddenly stuff that we did in a very naive way will need to become more efficient. And I think that's the next thing that we'll see. How we take, and skills, again, it's one direction... It's one great place to the direction. And I think that what we'll see is that we will improve on everything in everything's on the stock it's going to be improved for those use cases in a more efficient way than today because we are using the naive way. Make sense?

Rob Strechay

>> Totally agree. I think that's great. Well hey, thanks for coming on board. Really appreciate you coming over, and stopping in. I know we are both heading back to Boston pretty soon, so back to where it is cold. It's supposed to be cold. So, thanks again.

Idit Levine

>> Thank you so much. Appreciate it for having me.

Rob Strechay

>> And thank you for watching this episode. Stay tuned. We got a little bit more coming live from KubeCon CloudNativeCon, North America 2025. Stay tuned.