Insecure Agents — Episode 51: Catherine Jue, Co-Founder and CEO of Kernel
Podcast: https://insecureagents.com
Listen: https://open.spotify.com/episode/0FyhE1cMO2vC1dOru8slja
Transcript (auto-generated, lightly edited for clarity).

Allie Howe: All right, welcome to another episode of the Insecure Agents podcast. I'm your host, Allie Howe, and I'm super excited to be here at AI Engineer World's Fair day two. I'm with Catherine Zhu, co-founder and CEO of Kernel. Kernel is AI infrastructure for agents that are looking to browse the web. I'm super excited to get Catherine's perspective on that and also AI engineer here today. Catherine, would you like to introduce yourself?

Catherine Jue: Sure. Thanks so much for having me on the podcast, Allie. Really excited to chat. Yeah, I'm the co-founder and CEO of Kernel. We build crazy fast open source browser infrastructure for AI agencies, the internet. And that runs the gamut of solving the hard infrastructure challenges of running Chrome in the cloud at scale the security boundary challenges of that problem as well. And then the third piece that we s- aim to solve is the fragmentation of the internet. It is not a one size fits all solution. I mean, so we aim to abstract that away and make it super easy for a developer to give an AI agent a browser and it to do whatever it needs to do on the internet.

Allie Howe: Amazing. Super cool. What made you start Kernel? Like, what was the gap in the infrastructure that you saw that said, "Okay, like agents are missing this in order to be able to perform and access the web"? What, what was that moment where you're like, "Okay, we need this"?

Catherine Jue: Right. When my co-founder and our CTO Raph started working together computer use had just come out as a technology, and that as we started to play around with this, saw that it would open up a lot of new doors. And, this concept of a VLM being able to use the same human interfaces that you and I do on our personal computers felt really meaningful and compelling because it could solve a lot of challenges with the fact that most services in the, in the world do not have an API. But there are tons of challenges to solve in, in that problem space, not just the model inference being capable of doing that efficiently and accurately, but also the infrastructure challenges of hosting virtual desktops.

And then thirdly and, and where my co-founder's experience came into play is the identity story for agents. When an agent is operating on desktop machines using the rails that you and I do on our personal computers, a lot of questions open up about how does delegation happen, scopes and permissions. My co-founder was previously co-founder of the company Clever, which if you can think of as Okta for schools. He was really excited about solving the identity layer and that's what we're, we're solely focused on, on top of our browser infrastructure today.

Allie Howe: Amazing. Yeah. There's so many identity issues with like agents trying to access the web and currently, like I know there's a big like bot control program too, where they're being recognized as bots. And so I know... I don't know if you're familiar with like the web bot auth or whatever. Would love to tell our users like more about that, but would love to dig into more of the identity challenges you're seeing and how you solve those for our agents.

Catherine Jue: Yeah. So browser infrastructure is interesting because again, it's, it's an... It is both the rails that an agent can do anything on, on the internet if the end service provider does not have an API or an MCP server. Agents are technically capable of using th- that interface, which is really exciting. But in order to do that, you know, they have to ride the rails that we do with our human that, that were designed for human users When websites built an interface with a web op with authentication, you know, whether that is the website that powers kernel.sh or, you know, healthcare websites built 10, 20 years ago, a- an AI agent operating on behalf of someone else was simply not in these, the user journey story or an, or a user type that they ever accounted for.

And so that's why you h- you see these very, like, retrofitting, like shoehorning solutions that try to answer this question. There's a lot of unsolved challenges that the industry, kind of needs to address. We're aiming to do that by both enabling agents to use those, those rails, the browser infrastructure rails, and then layering on top of that, you know, identity, scopes, permissions, audit trails, logs things that are useful both for developer agents, end user consumers who may be delegating to an agent, and then ultimately for those end website services to know when agents are using their site, how they're using it, and if they want to enable or disallow an agent to do certain actions.

Allie Howe: Amazing. Yeah, it's-- agent delegation is so challenging because like you said, if you're a user trying to access the browser, I mean, you can do anything. You can, you know, go through your ch- shopping cart, buy something, check out. You can go back, you can navigate to like an admin panel. I mean, whatever users the permi-- like permissions the user has, agents are probably just inheriting those, like fully scoped today without this added au- authentication and identity layer that you're layering on top. Are you building all this yourself?

Catherine Jue: We're doing what we can. So, this kind of how I think about it is there are near-term solutions on what should we do on top of the existing rails, and then there's, you know, where does the industry need to head, and how can Kernel push that forward to where we need to go? So certainly in the current world where agents are just using the, the credentials delegated by a human, I mean, even that is not a trivial problem. It's like, how do those credentials get passed to that agent? Or how do they log in on behalf of an agent? Is consent collected anywhere? Where are those credentials...

Who-- Like, where are those credentials touching? How do we ensure LLMs don't see those credentials? That in, in and of itself is a, a pretty hairy problem. We've... In today's world, we've built a product called Managed Auth, which you can think of as Plaid for any websites. And so it is an SDK for our customer, the developer, to securely collect credentials from end users for a given sites, and then Kernel, the infrastructure provider, handles that login and re-authentication step such that an agent does not need to do that themselves, or a developer does not need to build that themselves. So that starts to solve like some of that initial challenge of even just secure credential handling on behalf of an agent.

It opens up a lot of future doors for us when we think about what will it look like for end websites to known agents distinguish a legitimate agent or authorized agent versus, you know, a DDoS attack, malicious bots, or just agents that they don't want on their site. We are kind of like solving the chicken and egg problem by solving the problems for the developers and the agent builders today helping them securely collect those credentials and store them and, and authenticate. And then it gives us, you know, the telemetry we need to then, when a website is ready for an agent to be using their website, be like, "Hey, we actually have the like audit trail of consent, authorization."

When we get into the world of scopes and permissions Kernel can play a role in there as well. That's kind of how we've, we've approached it so far.

Allie Howe: Did you see the enterprise Managed Auth release from Claude not too long ago?

Catherine Jue: No. I don't know which... I, I'm not sure if I know which one you're talking about.

Allie Howe: It's, it was so, so recent. Basically, they now allow you to use your IDP, so like Okta, to gate permissions to all of your different like MCP servers and, and tools you're connecting to. So there used to be this problem called like OAuth Islands, where users would use OAuth to get, you know, a, a token to log into some sort of application. But your IDP, your, your, you know, c- and your organization admin at a high level would have no insight into those OAuth tokens- Yeah... would not be able to revoke them, wouldn't have any insight to them. So it's a huge like underlying risk that's left over.

But the ID JAG token that just got- Released and also incorporated into sort of the backbone to Enterprise Manage Auth- Yeah allows for IDPs to be able to sort of like federate access to- Yeah... different MCP servers. So now it's actually the ex- Enterprise Manage Auth is an extension that's added into MCP servers, so you can use this new token ID JAG to, okay, if my ID PSA Okta is allowing me to access this MCP server, I can just use that token versus an OAuth grant- Yeah which is so much better for one- Yeah... 'cause like now the unlock there is that the developer only has to log in one time their IDP to access all their MCP servers.

Yep. And so it's a win for our developers, and it's also a win for security 'cause now they have visibility, and now they can revoke if they need to. What's the equivalent look like for browser agents? Yeah. 'Cause I feel like that gets really hard 'cause it's like, do you use like the ID jack token to say like allow list certain websites? Like, can you use that? Like would that... Does that feel helpful at all? Just like super interested to think like where, where's the industry going for browser agents for the same thing?

Catherine Jue: Yeah. It's it's something I think about a lot in part because again, like pre-LLM, what is the programmatic... Or pre, pre-browser agents, what is the programmatic way for an automation to access end services? It is OAuth, it is API keys. These are designed by end website services for programmatic use. Scopes are built into the, the protocol. Delegation is very obvious, audit trails. And then VLMs and browser agents kind of like blow that out of the water because now they're just-- they're able to really do anything on the interface that was really designed for, for human users. But it-- That also opens up a lot more opportunity because again, when you look at the, what is the surface area of all services in the world that have an API, it is like a fraction of all the things that you and I do on an end website service.

And so I'm really excited about browser agents being able to like unlock a lot more workflow automation, programmatic access in a way that doesn't require resources from those end websites. But then it is very tricky, and we can take a lot of the paradigms that exist at the OAuth IDP layer. And I think that that's really where the-- being in a third-party infrastructure is, is useful. We are not the agent itself. We are not the... You know, we can partner with end websites in, in designing the system that is both easy for a website to adopt while also kind of addressing all of the concerns that you mentioned around scopes and provisioning.

I think it's actually this identity layer and how we like make browser agents we give that to browser agents, is actually the main thing that holds browser agents back from being enterprise-ready. Because to your point, an enterprise needs, you know, federated access. It needs system controls for a single team to understand how different services are being used by an agent across different teams, and then being able to revoke that access. We have some ideas of how we intend to do that, kind of lay it on top of our managed auth product. One example is our browsers are fully sandboxed in, in a Chrome in a, in a, in a sandbox VM.

We can do some really interesting things like intelligent egress networking. If an agent is or isn't allowed to access a certain website or endpoint, that could be done at the, the networking layer. I think there's a lot of... We have some ideas, and we're kinda like early in this, but it's a big area of exploration for us because the It's, it's, it is one of the biggest challenges that holds browser agents back from adoption. The technology, the models themselves are quite capable nowadays.

Allie Howe: Yes. Yeah, no, that's so amazing. I actually, I opened up Worlds Fair yesterday, and I had to do an intro for the software factories track, and one of the key points I touched on or wanted to touch on was like, okay, hey, like, what's different between now and last year when Jeff Conley's Ralph Loof came out, that now we're just, like, at this inflection point for software factories, and it's because of a lot of different advancements across the industry. One of them is, like, the models have gotten better. Like you said, we've got better reasoning models. They're able to take on more sophisticated tasks now.

But also model vision has improved. So now models can verify work that they previously couldn't have before. And I think we see a lot of that in the browser use space 'cause now you can actually see, like, did it open the right webpage? Like, did the form actually complete? Did this thing submit? So How are you using those advancements at Kernel?

Catherine Jue: Yeah. We've seen, we've kinda seen the technological advancements progress even over the past 12 months. When we first started a year and a half ago, the majority of users on our platform, developers on our platform were mostly maybe doing some deterministic scripts, maybe they were having an LLM auto-generate deterministic scripts. But really this anchoring on like, I wanna know, using the DOM exactly, you know, with ID selectors, what it should act or not act upon. As the models have progressed and as vision-based models have improved as well as well as gotten, you know, not only more accurate, but cheaper, faster we now see our users like leaning more into the a- the truly agentic side, which is, "Hey, I'm I'm a developer building an agent.

Rather than like prescribing tools in very opinionated ways about how I should use the internet I will actually just give it the tool and give it the prompt." Yeah. And that concept is really at a point where we see production use cases at scale operating fairly non-determinant a- deterministic agents with tasks that I think is super exciting. On the concept of DOM versus vision, you know, there's, there's a lot of trade-offs there and y- the, the cool thing is that the models are getting better at both of those things, so they can both like generate DOM-based scripts much faster and on the fly and they're able to like take a screenshot, understand it on the fly and, and make a decision.

We see both of those mechanisms with customers on our platform today, and I think that will... both of them will only increase as, as the models improve as well.

Allie Howe: Yeah, that makes sense. Yeah, it's really exciting to see, like, and be able to react to all these new advancements and incorporate them into your product. I know that's like what makes now s- feel s- like such an inflection point for software factories and also just like for agents. I feel like so much is really coming together which is super exciting. And now we finally have AI security and engineering kind of in the same loop. Like we realize that we cannot have the engineering velocity that we want to- Yeah without security. And quickly people are realizing when they move from POC and trying to move into production, like, hey, wow, like these problems like identity and auth and delegation they matter because not only do we have users or agents acting on behalf of users, we also have agents acting on other behalf of, of other agents.

We've got these multi-hop delegation chains now where, you know, Agent B Got something that was delegated to it from agent A, and then agent B might be handing off to, like, say, a Kernel browser agent, which might be like agent C now. So, like, whose permissions do you use? The users all the way from agent A? Right. Like, or do you use a subset of, like, these permissions? Does, like, C have all just its own? Like what's-- what are your thoughts like there? Where does the industry need to go to be able to answer those questions?

Catherine Jue: Yeah. I mean, we, we-- Claude-- Anthropic just came out with something about this, around how Claude tags and delegation happens. And in a Slack thread, if Claude is in it, whose permissions is it using? Is it using a single user? Is it using the group's users? The answer is probably, you know, nuanced to the use case. I don't think it's a one-size-fits-all answer. In that example with Anthropic, I think where they landed is, like, Claude is its own identity. Sometimes... And that, that enables a lot of things. That enables it to potentially, like, you know, inherit the identity of one user when needed, but then have this audit trail of, "Oh, but it ultimately did pass through an agent with these, this user's permissions, this LDAP's permissions." again, it, it depends a lot on what the use case is, what's the interface in which you're interacting with the agents.

For Kernel and, and, and being an infrastructure provider, we really focus on, like, composable primitives that give our user, the developer, the tools they need to mix and match and build that, that delegation identity chain that they need. And so on the security front, for example, enabling... Again, rather than being fairly opinionated and being like, "Here's one box, black box for your agent, and you, you know, you j- you don't need to think about it," that actually is, like, not useful in this very messy world and where you have different use cases requiring agents to know or not know about different things. And so as we think about our product as an API and, and infrastructure primitives to surface to developers, we think a lot about, yeah, like, what are composable building blocks such that y- a developer can use Kernel and an agent with the right set of scopes, permissions, allow lists, networking requirements that they need.

And that's, that's kind of how we think of, like, our job in enabling people to build, like, verifiable, trusted chains of delegation.

Allie Howe: Something I was super excited that I read on- online was it said that one of your users is using browser use agents running on Kernel to register on government sites, 406 total processed forms, dozens of web agents running simultaneously, letting them process forms overnight. What does the infrastructure look like to be able to enable that? 'Cause, like, what's the orchestration pattern, the sandboxes that are needing? I mean, security aside, that's like an engineering, like, marvel to g- be able to get all of that to work, I bet.

Catherine Jue: Yeah. And One of, i-it's, it's both like a very exciting use case because it, it starts to highlight, you know, what is the power of a browser agent or, or how can the concept of an LLM have real world impact in a very meaningful way. From the infrastructure side of the house, that is, you know, the goal of Kernel is actually to take away that infrastructure for that user so they don't have to think about, "Okay, I need to scale my browser agent to tens, thousands, millions of browser sessions, even in parallel." That's a very meaty infrastructure challenge and something that we saw users were being held back by if they had to home-roll it.

I can think of another example who, when we started working with them, they were home-rolling and building a QA AI agent and they were like, "Yeah, it's totally fine, like we, we have this home-roll solution but we can't actually scale beyond X because then we'll need to invest a lot of infrastructure resources." They integrated Kernel, and it allowed them to actually start offering new products to their customers because they no longer had to think about that scalability challenge. Wow! And that was, that was really exciting and meaningful for us as, as the provider because it shows that like even just solving the scale challenge can unlock new use cases that were actually maybe holding them back that they didn't even quite realize when they didn't have... when they had to home-roll their solution.

For us, it's what we spend day and night doing, and so that looks like, you know, how do we manage Chromium, which is a h-very resource intensive application in a sandbox at scale, where a, any of our users could burst up to a, you know, thousands or a million kind of at a single time. Burst down as well. There's a lot of capacity challenge planning around that. And then how do you do that very cost effectively? That is like a non-trivial challenge, but it's also like the beauty of serverless is that we can like aggregate all the demand of, of people's different workloads and be able to like distribute that in an, in an economical way.

We build our browsers on Firecracker VMs, and Firecracker VMs are interesting because they enable a standby mode. So you can actually spin one up very, very quickly, you know, run your processes on it, and then actually like s-save its states to disk and then spin it back up whenever you need it. And that is a unique technology that enables us to be able to like scale up and down very, very cost effectively. One of the things that we kind of like hung our hat on very early, in the early days because it enables both really unique behaviors of our, of our browser VMs as well as gives us a lot of performance benefits as well.

Allie Howe: I wanted to go back to what you said earlier about being opinionated on the primitives-... that are in your platform. The beauty of infrastructure, which is I think what you just, like, mentioned essentially, is it allows developers to focus on just shipping, like the thing they're actually trying to build. They shouldn't have to spend hours on Auth. They shouldn't have to spend hours on scalability. So the beauty of that AI infrastructure is giving developers a lane to drive in where they're, they're doing what they're, like, meant to do. They're actually doing their specialty without getting lost in all this extra stuff. But what's the value in being a little bit opinionated for like the primitives when it comes to Auth, when it comes to scalability?

Sometimes developers think they know what they need, but maybe they don't. Where's the line there?

Catherine Jue: Yeah. It's a good question. At the infrastructure layer, it is certainly answering the questions like, you know, I think security is a critical component of being an infrastructure player. And so making the opinionated choice to not run Chrome and Docker, which was, you know, kind of the prior art to, to the world before browser agents and browser... Th-this world of like browser infrastructure really existed. We were one of the first players to really hang our hat on like, "No, no, we're gonna do this in like a fully sandboxed way," more technically difficult, but like security boundaries are gonna be really important for this, for this product over time.

At the identity layer, it's kind of like, it is to your point a fine balance because you want to both enable magical experiences for developers where they don't need to know, you know, have perfect information about how proxies work, how bot detection works, or, you know, what is the right way to implement a loop that understands a login page. But what I will say is over time, we have felt we have felt more-- we are more helpful to users when we give them more access to the primitives for them to, to, to decide what makes most sense for their agents. And that's largely because of the fact that, you know, agents can increasingly like do any task that you've given it, a coding task, for example.

It can do it quite quickly, quite efficiently. And so the... Where I generally see developer tools and infrastructure headed is, is less, very less black box. It is actually like exposing the right primitives that can be generalized across many use cases that can s-extend as the, you know, as our industry changes and evolves over time. We spent a lot of time thinking about those from primitives from that lens as opposed to again, like trying to p- overly, overly opinionate and package up too many things. We tend to find that they, they fall apart in six months because a harness that like was really opinionated and, and, and did a lot under the hood six months ago, no longer needed.

Claude can figure it out itself. It's really about giving it the right tools.

Allie Howe: Yes. Amazing. And like you said, those tools change constantly as different models improve, your harnesses improve. So yeah, no, I agree with that for sure. But yeah, it's a fine line between like giving developers what they think they need, but also just like the right primitives they can like drive fast and like move fast and ship fast, but it's also hard to probably pick out, like here's the primitives that we think are gonna be long lasting, that are just gonna drive all these advancements. What's happened like, you know, over the course of, you know, days, weeks, months, like they happen very quickly. So betting on the right primitives I think is something that's very difficult in today's industry.

Catherine Jue: We definitely think a lot about where are the model labs headed. Like what is, you know, what are the pain points today that developers have, and we certainly wanna help them today. And then we, and like in equal part think about like, well, where are the model labs headed? And you can kind of get some, some early si- insight by just listening to w- the way they talk about their world and what they want Claude or OpenAI to be really good about in the future. And then working back from there, it's like, again, how do we potentially like solve the, solve the world that we live in today, but not over-index on it because we want our primitives, our infrastructure to be enduring for whatever the next iteration of harness or, or kind of like capabilities of a, for an LLM is it's a fine balance.

Allie Howe: Yes. Yeah, for sure. I had a guest recently on the podcast Damian Shingleman, who's the VP of R&D at Okta-. But previously at Auth0. And he was talking to me about how at Auth0 they were trying to solve an issue where a person can spin up any number amount of agents. They might go on the web and try to, you know, post on social media or register for so- for something. And right now, like there's no way to tell like how many agents like belong to one person. It's hard to sort of like notice fraud when it comes to those patterns. So my understanding what they were like developing was this solution where you could tell if something was like an agent working on behalf of a user, but the user would stay like anonymous, to sort of help solve this.

Do you see like any misuse or fraud happening with like browser agents? Is that something you all have to think about at all? So like if you see that happening on your platforms, like, "Oh wow, this is a lot of costs, like we should probably shut this down."

Catherine Jue: Definitely. And that is, it is both the, the challenge of being in this space and also I think the opportunity for us to be a really great company is we are not just doing VM infrastructure, but like uniquely browsers are the portal to a lot... like any other service. And as a result, also a vector for, for malicious activity as well. You know, we've only been around for a year and a half, and it's something that we've hired security engineers around. A lot of infrastructure engineering with our founding team. And so from, from day one, we've had to think about this problem. And we kind of look at it from a variety of different ways, both in terms of monitoring existing workloads on our platform, cracking down when we see, when we see abuse.

And then of course is the, you know, we, we are very actively like excited about legitimate agents and enabling that them to be better distinguished from those bad actors on our platform and, and for the end website services. And then, you know, how do we do that for ourselves and for, for other end website owners, as well as how do we enable those end websites to know about the good legitimate agents? And, you know, how can we help them al- like distinguish and then better use better accept nuanced scopes and permissions for, for those agents on our platform as well?

Allie Howe: Yes, absolutely. Yeah, that makes sense for sure. What's something that you think is missing from the industry right now when it comes to identity or auth or infrastructure that would make your job easier if the industry were able to like come together on like this one thing or solve this next?

Catherine Jue: Yeah. It's funny when you look at the, you know, the early days of the web, and like what were the protocols that were built in they were really... You know, there was talk about two that never actually were implemented, and that, you know, many years later are the challenges that we now have to solve and what makes the browser agents so messy and complicated. The first is payments. And so there was, like, early talks about there being a protocol called 402 that meant, like when you hit that end-- when, when that response was returned, it meant you needed a payments in order to, to use that service.

And so we've started to see some of that happen in the, like, agentic commerce side of the house with, like, protocols from Coinbase, such as X402 coming out. Stripe and Shopify are doing a ton of work on enabling this over time. And then it is really the, the identity story because unlike other protocols Auth was not centralized. It, it was left as an exercise to the reader. Like- Yeah... Every website must build its own Auth standards and figure out exactly, you know, implements what-- how that nece-- that, those security boundaries exist and how scopes exist as well is why the web is so fragmented and why it, it is what, like, holds browser agents back quite a bit today.

So those are, like, the two that, that I see that I'm really excited about Kernel pushing forward on the latter but also really excited about the payment side of the house enabling agents to become more and more capable of really doing any task that you and I do. I think a lot of A lot more opportunity will, will unfold when agents are, have the right infrastructure and protocols to be able to, to do the tasks that they, they need.

Allie Howe: Yes. No, that makes sense for sure. Here at NCur agents, we think it's not only one single company, one vendor, one protocol is going to solve agent auth holistically. It takes a big community, it takes a big tent, everyone's getting together under that tent to figure this out. And one step towards that is we're running AAuth Night tonight. It's an event where the creator of OAuth is actually going to come in and tell us, you know, why OAuth doesn't work for agents anymore, why we need something new, which is this new protocol called AAuth that he's developing. Would love for Kernel to engage in that community, whether that's coming to that e-event or something in the future, just because I think we could all learn a lot from your experience, the challenges that you're facing and, you know, what does the industry need to provide to you to be able to give this sort of infrastructure to your users? 'cause you're right, like not every single website, not every single agent needs to go reinvent auth.

There should be some standard. And there's certainly best practices today, and I think the ID Jack story really helps with OAuth today. But ultimately, we are looking at new agent-native solutions for the future and we need people to come together and design and figure out what that is. So would love to see you all tonight or continue to have this conversation.

Catherine Jue: Yeah, definitely. I'm, I'm super excited about the event tonight. And yeah, it is very much a big tent moment. And it's also one that will evolve and I think emerge over time. When you look at the history of the web generally you don't find standards to be decided in one moment of time, then broadly adopted. It is very much a, a chicken and egg problem and, you know, the more that, you know, everyone involved can kind of just like keep nudging the world forward adopt o- like build open protocols build upon other people's protocols I think we're gonna get to where we need to go.

So yeah, excited about the future.

Allie Howe: Yes. Yeah, me too. What excites you about like Kernel? What's next? What are you building? What's on the roadmap?

Catherine Jue: Yeah. We've done a lot of work in the past nine months around becoming a verified agent on with WebAuthn programs, and so partner with Cloudflare, Vercel, Akamai to enable, enable agents on Kernel's platform to be verified and signed with our key. I'm really excited to extend that further and think about, hey, you know, to this, this question of which agents on our platform are malicious versus which ones are legitimate can we help builders of agents with our verified status to enable them to more easily be able to, you know, raise their hand as a legitimate agent on the web and then like pass through that through our infrastructure.

I think there's a lot of kind of like exciting partnership and product work there that we think will just enable developers on our platform to, to find be able to succeed on the web a lot more easily.

Allie Howe: Amazing. Yeah. So no, that makes sense. Super excited to continue to follow everything you're doing at Kernel, hear more about how this identity story evolves. Excited to keep in touch. Is there anything else you'd love to share with us today?

Catherine Jue: No, it's been a pleasure to, to be on on the pod. If you're interested in learning more about Kernel, we have a Twitter account. It's use Kernel. We're hiring a ton as well, and so if these problems excite anyone listening please reach out. Would love to chat with anyone especially if you're building agents in the infrastructure world, security world, and identity world.

Allie Howe: Amazing. And where people can be able to connect with you, LinkedIn, Twitter?

Catherine Jue: Twitter is kind of where the happening is use Kernel, and then my personal handle is J-U-E-C-D\_\_

Allie Howe: Amazing. Yeah. Okay. Well, thank you so much, Catherine, for coming on. Super excited to keep in touch, and thanks for the conversation.

Catherine Jue: Cheers.