Bart (00:00)
Hello and welcome to the Bright Signal Podcast. I'm here with my friend Raphael and we just sat together with Maurits Kapteyn.

Raphaël (00:08)
Moritz is the co-founder and CTO at Kyvvu. And today we are talking uh AI agents and more specifically, how do you stop them from doing something they shouldn't? And Kyvvu has the answer. It's a young Dutch startup founded in late 2025 that builds what they call an agent security kernel, which is basically a layer that sits inside an AI agent and checks.

every action before it happens. And so it's based on everything the agent has done before. Or as they put it, every agent has to ask before it acts. And behind Kyvvu, two experienced co-founders, Moritz Captain and Yeroun Haysen. And they already raised uh with Kyvvu one million uh seed round this summer uh and today we have the chance to interview Moritz who you will hear

really knows what he is talking about.

Bart (01:06)
Let's get to the interview.

Raphaël (01:08)
Yeah, let's get to the interview.

Raphaël (01:09)
Hi Moritz, thank you for joining and welcome to Brightenal. and I are really glad to have you on. Today we will talk about AI agents and security and about your company Kyvvu that you co-founded. You are a university professor, you write papers, also books, but at the same time you're an entrepreneur. You founded three companies, exited two. That's kind of impressive.

But before we get into that, you maybe start by introducing yourself and explaining a bit what you were doing before starting Kyvvu?

Maurits Kaptein (01:47)
Yeah, awesome. Thanks for the invite. super happy to have this conversation. a bit about myself. So uh I'm well you introduced it very properly, Max Kaptein. I'm still a full professor at the University of Eindhoven. I'm under mathematics and computer science department. So I have in part an academic career. That career

Started at some point like working in my PhD, working on kind of reinforcement learning type problems and then and a bunch of different like statistical problems. So I've I've always been interested kind of in this intersection of statistical decision making, but then also like how do you compute these models? And obviously that like turned into what's what's happening in machine learning, what's happening in AI, like all of these things logically followed. but

One of the things that I did in my PhD, which is over fifteen years ago by now, because I'm getting old, um, was basically develop at the time we were kind of

Yeah, the like a a a bunch of internet and e-commerce and all of this was popping up and people start thinking about how do we personalize web pages massively basically. So so how how do I like potentially give you a different offering than somebody else? Um which of obviously kind of relies on like well how how do you then compute kind of based on the information that you have what the next best page would be, basically. so that's one of the things that I kind of worked on in my PhD, and that immediately

because of its clear application spiraled into being a company. with that company very early on, like Booking.com in the Netherlands was one of our customers. They were quite

Bart (03:18)
Cool.

Maurits Kaptein (03:18)
a bit smaller at the time, by the way, because they were obviously like a Dutch small company at some point. uh But I think we made some some impact on like how this is actually done, like how how personalization is done on the web.

Which I thought was very cool. But kind of after a year in the company, I decided to focus on my academic career. Obviously, like if you want to become a professor, there's there's some work work involved. So I spent time at a university, various problems that I worked on, mostly like a problem called contextual multi-arm pendits is something that I've worked on for a long time. happy to talk about that more. Uh and at some point moved a bit more to like causal inference type problems. but around

Eight years ago I started a company which basically like

Again to like the first one was like, the technology is moving in in a specific direction, there's new problems coming up, which means you can think about cool solutions for new problems basically. The first one was like, How do you deal with this massive click stream data if you want to like personalize in real time? It's a cool computational problem. This one was slightly different. We saw all these machine learning models, and obviously, like deep learning models start to pop up, and now suddenly I think the first wave, at least for me consciously, was like all

All

of these vision models really starting to work. So it used to be, even when I was doing my PC, like I had colleagues that were really, really trying to make like how do I recognize license plates on images basically. And then at some point, by the time we got a hang of like large data sets that were annotated and kind of deep learning, we'd start to just build these models, train them, like with very little, like you know.

Engineering, just train them and they work very well, basically. And I think around that time it became clear like these models are coming up.

Well, at some point, if you can train them very easily, you want to use them. And if you want to use them, you have to deploy them somewhere. So you have to do managed deployment. And the company, like initially, we just started at like how do you deploy these models like efficiently? So kind of low-level, how do you compute some of these things efficiently? But because the way we were building on this, it also made it very easy for us to pour to specific edge devices that had like these specific

Bart (05:30)
Okay.

Maurits Kaptein (05:31)
chips and and like small chipsets and whatever. So so it actually ended up being a

Product that kind of managed model deployment to constrained edge devices basically. That company, yeah, we we grew. I was the CEO of the company at the Athletic Jur during all of its time basically. And about three years ago by now, it was sold to a company called Network Optics. Network Optics builds well software for security cameras. So in some ways that that

Bart (05:58)
Okay.

Maurits Kaptein (05:59)
was a really good fit. Spent some time at Network Optics as well. It's kind of cool to like travel the world.

see you see where the whole product went because now suddenly we had a bunch of distribution of the whole thing so that was cool but then at some point yeah I kind of yeah in some ways had enough of working for a bus sorry to just put it that way but like the decided to leave

and decided to start thinking about new things and kind of one of the things I was

Obviously coming up, sorry for the long introduction, but I'm getting to Kyvvu and then you have the whole story. Um but

Bart (06:27)
No, no, it's very interesting.

Maurits Kaptein (06:29)
basically like what was coming up at the time, obviously, like everyone had the chat GPT moment. we can chat with these things, but it became I think everyone playing with these things, it was clear, like, if I can generate this type of like freeform reasonable text basically, well then maybe I can also just generate a plan. Like, maybe I can ask this model, like if you would do a task, like

What steps would you go through? would you first read an email or go searching on the web or write this database query or whatever? And like, these models gave pretty good plans. Like obviously it it all became better over the last few years, but the the the idea that you could have not just content, but really like actions being produced by such a model, I think that was kind of clear for everyone playing with them very early on. so we started.

I do think it's as many people in the space, building our own harnesses. It's like basically a bit of code that just executed whatever this plan was. And then

I guess by now, like around two years ago or one and a half, the you know, like length chain, like there's more and more of these like harnesses that were doing well.

Bart (07:38)
Hmm.

Maurits Kaptein (07:39)
It's easier to build these agents that were kind of these somewhat open end, like these loops. People start talking about loop engineering. I th I don't know the the bus terms change all the time, but it was basically like, now, what if I give this a like this LLM this list of all the things that it can do? I give it a task and I have it decide itself, like how it's gonna go about. And as long as I just automate the X.

Execution of the plans that come out. Now this thing can basically carry out this whole thing by itself. Which is cool until like we had our first accidents building them, accidents being like you know, secrets in a secrets file being shared just on a public GitHub, deleted databases. We've seen a bunch of like weird weird stuff basically. and to be honest, this is kind of what

Where Kyvvu came from because in the end it was like, if we are using these things and we're

We're giving these agents more and more space to act, basically, more and more tools, more and more freedom. And I I I in some ways I think I don't know if freedom is the right word. So it's st it's hard for us to talk about these things without human terms. Yeah, autonomy

Bart (08:38)
More autonomy.

Maurits Kaptein (08:41)
is even more. But we're giving them more autonomy, which means now they can do all these free, like open-ended tasks, and and it's kind of amazing what you can do. But with that autonomy, obviously also come quite big risks. And and it's still fundamentally the case that like these LLMs

Behave pretty much non-deterministically, they make mistakes. Like a few years ago we thought it was a hallucination. Now it's potentially a wrong plan. And it it is potentially like a deleted database, basically, if you just faithfully execute what's happening. So with Kibu, we started building for how do we make sure these things behave? Well, we can dig into this some of the details, but started there a bit over a year ago, you know, and and by now we have

Yeah, well it's a fast growing team. So so that that's where we are. And

Bart (09:26)
Nice.

Maurits Kaptein (09:26)
some background about myself. There we go.

Bart (09:29)
Nice story. I think the autonomy part versus what can go wrong and security,

it's

Maurits Kaptein (09:35)
Yeah.

Bart (09:36)
very recognizable. Also, if I speak from my own experience, like it's very, it's a bit, there's a bit of a tension between in the sense that like I have the feeling the more autonomy or the more access I give my agents, the better the performances.

Maurits Kaptein (09:48)
Yeah.

Bart (09:49)
But at the same time, like you also increased risk, right?

Maurits Kaptein (09:54)
Yeah, it's it's it's it's a weird it's a weird like you know, almost

Paradox in some ways, and and in

Bart (10:01)
Yeah.

Maurits Kaptein (10:01)
some ways I wouldn't even know how this is gonna play out. So one of the things that I I thought was somewhat interesting is obviously the people that are training these models and especially like the big frontier labs are trying to make these models such that they give reasonable plans most of the

Bart (10:16)
Mm-hmm.

Maurits Kaptein (10:17)
time. Like, that is like I would say this you can you know munch all of that under the work of alignment, but basically like the output of the models should be reasonable.

Which means like weird errors like delete a database, sorry for this stupid example, but let's carry it on forward, are less likely. On the other hand, these models are becoming better, and we're giving them like more tools and we're giving them more tasks. And so, in some ways, like with more capability and alignment, like the risks potentially of individual errors or something decrease, but

the number of tests that we give and the consequences of these tests become bigger. So I'm not

Bart (10:57)
is way more.

Maurits Kaptein (10:59)
even sure whether like we're decreasing our expected value in some sense

Bart (11:03)
Hmm.

Maurits Kaptein (11:03)
of what we're doing. but it's it's it's interesting to just see. It's also just interesting to be in this space that's moving so fast, basically. We'll we'll we'll see where it goes. Yeah.

Bart (11:13)
Maybe before we dive into what you

Maurits Kaptein (11:14)
Yeah.

Bart (11:15)
do exactly with Kyvvu you started in January of this year, right? Of 2026. um

Maurits Kaptein (11:20)
Yep, pretty much. Yeah.

Bart (11:23)
Tell me a bit about the founding team and the current team that you have.

Maurits Kaptein (11:28)
Cool. yeah, so the the founding team in some other ways initially like I was working on this I was working on like building agents together with a a guy called Joris Jansen who's at at On and like a a very large

The Japanese company, and we we ended up writing a book together, uh AI Agents at Work, which we wrote like we released it in November last year. Um, and I think it's still kind of timely, so it's kind of interesting how how that evolved basically. but that was to me like a lot of work that we kind of did on building these things, provided the groundwork for for my thinking about what Kyvvu was. To be honest, like my

Bart (12:07)
How interesting.

Maurits Kaptein (12:09)
like the the the actual inception of Kyvvu.

Was myself going to the Dutch Chamber of Commerce by myself. So at the time, which is really starting there. But quite quickly, I had a number of people basically like advising or just helping out. And I think this is something that I've done with earlier startups as well, to have just some network of people that care about what you're building, that have their weights to specific customers.

There was also our first launching cons customer as well came through this kind of network. And a number of those people are still involved. I don't know if you need any of the names, but that like Boss, Coco, Henry, like there's there's there's right now effectively five people that just have a role like this that just help help out

Bart (12:49)
Okay.

Maurits Kaptein (12:50)
in various ways. but quite quickly, I think within like one or two months.

I was I was talking to Jeroen Geissen who who was actually a a board member of mine in the previous company. So in the previous company

Bart (13:03)
Hi.

Maurits Kaptein (13:04)
we had like a VC, like we had VC money, the VC put like a board man, like somebody on our board, and Jeroen ended up being being that guy basically that was on our board. So so

Bart (13:14)
There you go.

Maurits Kaptein (13:15)
I ended up working with Jeroen for five years in the previous company, and he was like super intram instrumental in the whole s you know the the acquisition in the end.

So we worked quite closely together. He himself has

history he founded a company called Meteorological at some point which did nothing with the weather so it's a poorly chosen name but but but

Bart (13:35)
Ha.

Raphaël (13:35)
You

Maurits Kaptein (13:37)
it was but but it's like it's an interesting like Dutch company story. It was acquired by Comcast I think eight years ago

Bart (13:44)
Okay.

Maurits Kaptein (13:44)
or something. But they basically built the plumbing for your smart TV. So if you like click on the you know Netflix or some of the other logos

Bart (13:54)
Yeah.

Maurits Kaptein (13:54)
basically well obviously like you you got

Yeah, that's an app that's finding streams

Bart (13:59)
Mm-hmm.

Maurits Kaptein (14:00)
somewhere, so there's a bunch of stuff happening behind the scenes. and this is what they built and and took global. It's it's like one of these, I think,

Bart (14:06)
interesting.

Maurits Kaptein (14:07)
very cool technical companies that nobody has necessarily heard of, but that actually played a huge role in kind of like that that whole scene. And obviously, like joining Comcast is one of like the largest or I don't know if this is the largest, but one of the largest cable providers in the in the world, I guess.

Bart (14:20)
Yeah.

Maurits Kaptein (14:21)
so so he has quite some background and he joined like we we b basically co-founded Kyvvu.

Like what's in a name, like everything it's always a bit more messy, basically. Um,

Bart (14:30)
Yeah, yeah, sure. Yeah, yeah.

Maurits Kaptein (14:31)
so so then Jeroen and I, and like because of Jeroen's own background in his previous company, he he he had been like a founder and CEO throughout all of that journey. And I am I've a bit more of a tech background, so I decided to take like the CT overall. Although I to be honest, I think the naming in startups is always a bit, yeah, you know, you do you do whatever basically we we and and and

Bart (14:52)
Yeah, it's the, in startup phase, the rules are always

a bit, a bit vague, right?

Maurits Kaptein (14:57)
And and it's whatever

what whatever nobody else in the team is doing and that this kind of stuff. I I mean sure, like I I in some ways have a final say in where we take our products per perhaps and he's doing a bit more of our commercial development. But that's that's

Bart (15:08)
Yeah. Okay.

Maurits Kaptein (15:10)
how it started. And and then we hired Fadim, so like a lead engineer on our side. We quite recently hired Sven who's doing like our revenue operations right now. So it's starting to starting to grow and we're we're still

Still hiring right now, so

Bart (15:27)
Okay, interesting.

And you're in the, quote unquote, agent security space, which is one hyped space, also very broad space, right? Like it goes from,

Maurits Kaptein (15:35)
Yes. Hmm.

Bart (15:38)
I don't know, agent identities, authorization, input-output protection, monitoring,

Maurits Kaptein (15:41)
Yeah. Yeah.

Bart (15:45)
behavioral policies is probably a bit the space where

Maurits Kaptein (15:46)
Yeah. Yeah.

Bart (15:48)
Kyvvu is in. Can you explain to us a bit like what's...

Where in that space we can find the key what you're offering to clients is.

Maurits Kaptein (15:54)
Yeah.

Very good question. and like the only thing the hype, like when we started it wasn't hype yet. I guess it's hype now, the whole thing, agent

Bart (16:02)
Yeah, yeah.

Maurits Kaptein (16:03)
security, it's it's it's it's just kinda interesting how within the course of like nine months

Also this our customer conversations and stuff changed from from sometimes like what's an agent to to start with

Bart (16:16)
Yeah.

Maurits Kaptein (16:16)
to like there's risks of running your own agents to now a lot of people caring about like security of their agents, which is

I would say both a blessing because it makes like now these product categories are starting to emerge. It's also a bit of a a curse in the sense that well actually in the sense of what you were just saying that there's like there's so many things that you could munch under

Bart (16:38)
Hmm.

Maurits Kaptein (16:39)
agent security. If you d if you just think about agent security

By the way, setting aside the whole field of securing yourself against agent attacks, which is like a very interesting field as well, but that's that's not what we're in.

Bart (16:51)
Yeah, fair point here.

Maurits Kaptein (16:54)
But the the the kind of if you think about it as like

making sure my agent does pretty much what I want it to do, let's just put

Bart (17:01)
Mm-hmm.

Maurits Kaptein (17:01)
it that way. It's a very vague notion of of what agent security would be. Then there's just so many approaches that are starting to emerge and and some of them are super sensical. And I would say like from our point of of Kyvvu, like we do, yeah, we do a very well defined set of things. but there's definitely other things that you want to do in like a mature, like agentic

company. some however are also a bit like I would say dubious whether they actually do what they want what what you what you imagine them to do. But I think like by and large you could like the whole field falls falls apart in a few different directions. And so um one is like you can think about the the content basically this the text that's going like into and out of the LM

Bart (17:49)
Mm-hmm.

Maurits Kaptein (17:49)
and and kind of the the actions basically and and and those two two

It's are kind of interestingly mixed in agents in a way, because an LLM is generating both like if you ask an agent to write an email, it's writing like the content of the email, and it's also using like it's saying I want to use a tool to send this email.

Bart (18:10)
Yeah.

Maurits Kaptein (18:10)
So it's also generating the action. But I think on the content side, so what's in that email, there's a lot of people that are trying to work on

Which I would call guardrails, but it's it's a bit of an overloaded term because people are using it for everything,

Bart (18:26)
Mm-hmm.

Maurits Kaptein (18:26)
like to make agents secure. But it's really it's checks of that content, which might be like, there's no credit card number in this contact, it might be the the tone of voice, it might be all kinds of things that you wanna kind of detect on what's actually happening in the content. and methods kind of range from

human oversight, I guess, having humans check these

Bart (18:45)
Mm-hmm.

Maurits Kaptein (18:45)
things, to like other agents checking the content of this age, so LLMs checking LLMs to like classification models, which I would say is even as far as I understand, something that like some of the Frontiers models, like labs actually are doing for their alignment work. It's like having a bunch of like just more traditional ML like classification models run after the the content is

Bart (19:07)
Yeah.

Maurits Kaptein (19:07)
produced basically. To

regular expressions, whatever, like but there's this whole suite of people that are focusing on the content basically. I would also like s think about people that that control like for example

There's there's a number of people that are working on like if the an agent writes something to memory and retrieves something from memory, like what was actually written out and what was retrieved. Because if you want to kind of know what this thing knows, you have to have some some feel of what's happening there. It's also check with the content. then the other side of it is like all the actions that the agent is taking. and there I think the whole stack is kind of starts with

Identity, which is something we should discuss, by the way, because I think identity is this somewhat overrated term in this whole space

Bart (19:51)
Mm-hmm.

Maurits Kaptein (19:52)
as it's starting to emerge. But at the very least, like access control. We used to have like rule-based access control that or role-based access control. now you see people working on kind of short-lived tokens for specific tools such that your agent has you know access to do something, but it's revoked quite quickly, so it cannot be abused, and blah blah. So there's like identity access control.

control are working on that space. There's people working very low in the stack, which is like sandboxing. So one of the

One of the potentially annoying thing about agents, if you if you really think about like what these things could do, is they could change themselves, basically, which makes like if an agent can rewrite its own code, then potentially some of the checks that you would write into their code can just be written out. so maybe you want to start this in kind of a controlled process where at least the the files that make up the agent harness cannot be altered. there's even like there's even hardware solutions there, basically, that people are focusing on.

Bart (20:51)
I think they're

NVIDIA, not that long ago came out with their own sandbox. What's it called again? Open Shell,

Maurits Kaptein (20:55)
Yeah, true. Yeah.

Bart (20:58)
right?

Maurits Kaptein (20:59)
so they have open shell, they also had some some noise about even hardware checks there

Bart (21:03)
Okay.

Maurits Kaptein (21:03)
on on this thing. there's also a project project called Shady, which I think is a somewhat poorly chosen name. but but but it's uh uh it it's like it's a Linux foundation project. It's they they have a a a very good sandbox as well, I would say. And and there's more sandboxes and it's like if you run whatever cloud code on your own machine, it will run

Partly sandbox with depending on which system. Uh yeah.

Bart (21:26)
But it's quite like

high level secure in a sense that you can say you can, you're allowed to do these type of file system changes. You're allowed to do this type of network access, but it's not really in the agent logic or in your application logic, right?

Maurits Kaptein (21:35)
Yeah. Yeah, maybe this network. Yeah. Yeah.

No, no. So so then we go to like kind of where Kyvvu is, basically. so we're we're not checking the the content. We're we're really thinking about the actions an agent has taken. And really one of the core

backdrops or beliefs or something, I don't know how to best call it, that that's you know, that sparked what we what we're building is the fact that individual access to individual tools might be

Okay. If you have a open-ended agent that can do many tasks. if you have just something that's just as a consumer, something that you know would be able to edit your pictures or something, cool. Uh and be able to send your emails, great. but maybe you don't always want it to send like randomly send your e like your pictures to whatever your boss or something. So some combinations of things might not necessarily be great, despite the fact that for that

agent identity, yeah, it should have like access to these tools tools.

Bart (22:37)
Mm-hmm.

Maurits Kaptein (22:38)
So then you get to like, there's the sequence of actions is actually something that we also want to cover. Saying explicitly also, like we wanna check content, we wanna have these sample, like we wanna do all these things, but the sequence of actions is something we wanna cover.

By the way, obviously, if you think about access control, it's a special case because that's like you know, a sequence of action of one, or just that it's

Bart (23:01)
Mm-hmm.

Maurits Kaptein (23:02)
a rule on the sequence of actions that does not depend on the history, basically. That's a way of thinking about it. but with Kyvvu, what we build, and I think like a way of thinking about it, it's what's called in like the traditional security literature, a reference monitor. So it's kind of this process that sits next to a process that sees every step that that process is taking, and allows

You need to specify, in our case, deterministic rules over what this thing is doing. and

We do this in a way, I don't know if you want to g go into the this much like technical detail, but but I do think it's kind of even in the way we like talk about what we do as a as a product, we we do this by integrating directly into the harness, so into the thing that's actually taking the actions. Why? Because then if we integrate properly and if we're running in a sandbox, we can actually give effectively guarantees that you know

Bart (23:57)
Yep.

Maurits Kaptein (23:57)
specific combination of things that you're you

you you say should not happen will actually not happen. There's also

Bart (24:03)
you can really interrupt

the application logic.

Maurits Kaptein (24:05)
You can you can

interrupt the application logic and all your events come from the application logic. They do not come from the LLM. And this is a very specific choice on our end. One is for the general kind of idea that the LLM is basically the untrusted thing in the whole in like in the whole setup, basically. But the other one is also, and this is where I think like, you know, I don't wanna like bitch too much about what other people are doing, but but kind of in shape. Some people are kind of sniffing the

wire between the LLM and the harness. Because obviously, the action plan comes from the LLM. So you have this plan, and if you just read that content, if if you were reading content anyway because you were like a content-wise guardrail, you can now also read the plan that's being submitted and potentially like intervene in that plan. What I find kind of at least theoretically somewhat tricky about this is that it implies that you know

Exactly how the plan is interpreted by the harness that receives it. And this is potentially tricky and potentially something that can be exploited. Basically, you're kind of looking at the intended actions, you're not really looking at what's actually being done. And this is why we chose to just focus fully on integration with the harness, because we want to actually just just focus on what's actually being done, basically.

Bart (25:25)
Yeah.

Maurits Kaptein (25:26)
There's other other things that people are doing. One of the I would say interesting parts obviously of our product is that by the time you're in this harness, you're seeing what's actually being done. You see like you have a set of policies that you can define yourself. We can we can talk more about the details there. but it also gives you the whole trace of all the steps that have been done. And you can guarantee which policies ran at which point. So now you also have your audit log m and these kind of things basically as a yeah.

Bart (25:54)
Maybe to

try to tie that to your example of the

pictures. So

Maurits Kaptein (25:57)
Yeah. Mm?

Bart (25:59)
you have this trace of everything that happens and you can define certain policies. So that means that like, if, I don't know, the agent took one of my personal pictures and I now in a next step um want to send it to an image editing tool.

for myself, like it will probably pass. But you can

Maurits Kaptein (26:16)
Yeah. Yeah. Yeah.

Bart (26:19)
say like, you have a policy, like if you took some of my personal depictions and I want to send it to my boss, then you need to first

Maurits Kaptein (26:24)
Yeah.

Bart (26:24)
ask me for approval, something like that, right?

Maurits Kaptein (26:26)
Yes. Yeah, this is actually a good example. And I think the way we often think about it is slightly more general. And so if you think about like more like most of our customers are like larger companies, we we don't really have a consumer product, but it would be like a a read of like classified internal data, which is what your pictures would be. Um which from that moment in time within the task, which is kind of like the memory scope of this whole thing.

Any right to an outside source will not be available anymore to the agent. Like that can be that's a potentially a blanket rule. So now you don't have to say, like, if you do a picture, because obviously, one of the kind of sorry for this the little digression, but one of the abstract ways of thinking about this is like you set a task, and like an agent takes a bunch of steps to complete that task. So all the steps it had available was kind of a tree like all the potential pets it could do, and really what you

Want is to exclude all of the pets that basically you find infeasible. I think that's

Bart (27:27)
Mm-hmm.

Maurits Kaptein (27:28)
like a much more defined way of saying what's agent security. Well, it's it's it takes a you know a a pet to do a task, and I only want it to do specific pads, not the ones that I find infeasible. so we're often looking for rules that are

Yeah, slightly more general. So so so kind of like

Bart (27:43)
Yeah.

Maurits Kaptein (27:44)
internal reads should not should never be followed by right to the outside world. However, and I think this is a really cool example that you kinda highlighted by your your question, is like, well, maybe maybe we should allow for some rights, but only after

Human approval, or maybe after like approval from some other agent, like after some process that approves it. Now, the way like we would operate, we as as Kyvvu on in such a process is effectively that we would set this rule that you do an internal read, you want to write to the outside world, you have to have human approval, which means like you have to pass this step in your in your task. And as long as you haven't, the agent just like any kind of write to the outside world would be blocked. The moment it

Happens, which means effectively that the harness logs human approval. Now that's in the taskpad, basically. So now we know it has been approved, or it it may that we can even enforce like the approval had to happen in our previous step, like nothing can happen in between. Okay, now it can send basically. which also kind of I think it's kind of an interesting way of where I would see.

Kind of content-wise checks and what Kifu does really interacts because, like, if you I mean just you want to send an email and you want to make sure that there's no like PII in the like personal identifiable information, well, likely you'll have some classification model to classify whether there's PII in this thing. That classification model likely has properties, like it has some accuracy, some version, blah blah blah.

Now, what you could do do with Kyvvu is for all of the agents in your organization, say any email that's ever sent always has to first go through this specific you know version of my gate before it can be sent out, basically.

Bart (29:34)
Yeah, uh

Raphaël (29:37)
to be a bit more concrete, I'm

Maurits Kaptein (29:39)
Yeah.

Raphaël (29:39)
sure there are plenty of use cases we cannot even imagine. to be even more concrete, can you give us some example of

Maurits Kaptein (29:42)
Shit, by the way. I thought this was quite quite concrete already, but my yeah. Yeah.

Raphaël (29:49)
customers that you have, maybe,

Maurits Kaptein (29:53)
Yeah.

Raphaël (29:53)
and their different use cases, and how did you solve their issue, basically?

Maurits Kaptein (30:00)
Yeah, it could do a very specific one. there there's a there's obviously like a bit more like

you know uh stuff happening around because now you have to start thinking about these actual tool calls and and like database reads where do you get your classification a lot of larger companies already have data source classification and and these kind of things um uh but I'll do the kind of easy story which is we're working with a a Dutch health insurer and uh this it's something I didn't know before we started working with them but

well let me first to the context like they're they're they're working towards building a customer service agent

Bart (30:38)
Mm-hmm.

Maurits Kaptein (30:39)
themselves. and really what they're doing right now, which are two things also kind of interesting, or what they did before like we started working with them, what we're doing with them isn't live yet, it's they're still like under development. but what they do have live is kind of a scripted workflow, basically. So so basically if you think about these pets, they kind of hard-coded the pets. So now, yeah.

Thing does a few LLM calls, but it's it's really like it's really confined what it can do basically. They would like to build this agent to be much more generic so it doesn't have to classify first what am I trying to do, but it can kind of like reason and have a conversation and answer questions. It should be able to answer questions about.

your prior claims. So if you're if you're a customer, obviously like there's logic on how you log in and how you identify

Bart (31:25)
Hmm.

Maurits Kaptein (31:26)
yourself as a customer, blah blah blah. But the high-level story is if you're a customer, you should be able to ask for your claims. Now that agent should have access to your individual insurance claims.

It should also be able, amongst other things, to answer questions about the the insurance packages. Like in the in the Netherlands you have like your your generic health insurance and then you can like top it up with like dental and and these kind of things basically. and I was told, like may maybe maybe maybe you should have looked this up, but I believed them. Like I was told that you cannot advertise your specific health, like you know, these premium packages.

Using somebody's individual claims. So you cannot or you're

Bart (32:06)
Okay.

Maurits Kaptein (32:07)
not allowed to effectively say, you've been to the dentist twice this year, you should take this dental package. uh

Raphaël (32:13)
All right.

Maurits Kaptein (32:14)
Which effectively comes down to this agent that has like in its core, it has two tool calls that it can do. It has more, like it's a more complex thing in the end. But like the core of this argument is it has two tool calls. And as long as it does one of those and generates a response, it's fine.

The other one it generates a response, it's fine. If it does both and generates a

Bart (32:36)
Hmm.

Maurits Kaptein (32:37)
response, potentially the content's actually illegal, basically, with you know whatever ramifications that follow from it. so that's an action that we block, or really like the way we're working with it now, it's like, that's one that would be parked for human approval. So the agent kind of like partly stops the tasks toward the user and like the message sent for human approval, blah blah. That's how how do you put it in the workflow? But I think this kind of

Hopefully somewhat illustrates like a a type of very specific pet dependent rule basically.

Raphaël (33:02)
Yeah, that's interesting.

on that, how does Kyvvu know that the output, the answer, actually used both sources, like the personal claim on the one side

Maurits Kaptein (33:15)
Ah so

Raphaël (33:16)
and the insurance product on the one side, and not just having touched them?

Maurits Kaptein (33:21)
Uh interestingly, we don't. so the so the whole point is we know it was able to do this. So we know by the time it's queried those two things and then generated a response using an LLM, that that's potentially tricky, basically. if you would want to know whether that specific answer is tricky, yeah, you would have to check the content, basically. so maybe this is where, and in some ways, like I would say by enforcing human approval at that.

Step, that's basically an enforcement of a check of the content, which is not necessary if it's only based on one or if it's only based on the other. Again, by virtue of those readings.

The moment both are red, this is potentially dangerous. And this is it's kind of like I think it's an interesting question because it's kind of this separation between like the action sequence versus the content that's being j being generated, basically. And I think the two interplay and and we focus in some way solely on the action sequence and and what those steps are, and that's something that we can then like give guarantees on. The whole content spiel, however I say it, is something that

In many ways I think you add up it like add to it. I think the human approval thing will not be scalable, so likely it will be some other way of like some you know, just some network like recognizing this based on a benchmark, that you have some accuracy, whatever it's for this specific task. I think that's where we will be going in the future. Um but

Bart (34:50)
Mm-hmm.

Maurits Kaptein (34:50)
what Kifu does is just ensure that like these exact steps are possible, are combinable, are not combinable. That's the type of rules.

Bart (35:00)
interesting. And it's also very interesting that you have this, it's a very deterministic way of enforcing a policy, right? for these and these

scenarios, you know that we will not allow to pass through this gate or ask for approval or whatever.

Maurits Kaptein (35:13)
Yes, well and I and I think this also just comes down to like two things that I think are kind of challenging if you're building your own agent. Like s some of them we don't necessarily like do anything about with Kyvvu, but like if you're building your own agents and you wanna know whether well I think it's always a reasonable question to ask like, if I have a new version of this agent, does it work better than my old version?

Bart (35:36)
Hmm.

Maurits Kaptein (35:37)
Well, that means like that task is hard compared to what normal software was. One, because of like the super open-ended tasks that we're giving. So it it becomes hard to measure if like it now does better on one task, but w what about all the other tasks basically? So so now you you have to start having like kind of a a a very broad test suite of things.

But that's compounded by the fact that these things are non-deterministic, which means that the exact same inputs might lead to a different output. Which means now like it's a probability game. Instead of instead of running through all your test cases, now you actually have to run through them in a way like multiple times to kind of know the probabilities of of them doing these things well. And

What we're trying to do, like when designing Kyvvu, was in a way, like, if this this this whole probabilistic side of it is one of the very tricky things, we can have all kinds of checks, like another agent uh basically checking this agent. It will change probabilities of things going wrong. All the alignment work that the large labs are doing, it changes probabilities of things going wrong.

Bart (36:46)
Exactly.

Maurits Kaptein (36:47)
But we looked at it like, what is this? This baseline set of things that you can really.

And like guarantee always happen

Bart (36:54)
Yeah.

Maurits Kaptein (36:55)
basically. And and this is also how we think about in the end. I mean, maybe even maybe this is kind of like a bit of a weird state, but where would like Kifu to be at some point? It's like, this is this base set of rules to like think about it as a constitution or something that

Bart (37:09)
Mm-hmm.

Maurits Kaptein (37:10)
agents in your company or maybe even broader should always comply to. Like this is the like such

Bart (37:16)
Yeah.

Maurits Kaptein (37:16)
such like the blanket, you know.

base sets and then on top of that you might have all kinds of desirable things that are all like much more probabilistic. It's good. Work on those. We provide this like baseline set of things that will definitely not happen.

Bart (37:32)
Yeah, could also like the way that you express it also very easily like extended to things like guard trails because like a gate, a deterministic gate can be

Maurits Kaptein (37:41)
Yeah.

Bart (37:42)
like the example that you gave, push this first through this PII removal set to

Maurits Kaptein (37:47)
Yeah, yeah.

Bart (37:48)
this credit card number removal set and only then is it allowed to pass, right?

Maurits Kaptein (37:51)
Yeah, yeah, no no so so I think

there's a lot of interaction between these things basically. And it's like all of your probabilistic checks the fact that they actually ran. we can enforce this basically. And and and that's and this is

to some people maybe this sounds a bit silly because they would say, Well, I you just generate a message and then you always pass it through your PII check before you send it. Yeah, well if that's what you do, if effectively you're scripting a part of the pet. Basically you're hard coding a pet part of the pet. The where we are going, it seems like, is that we say, Well, you can draft messages, you have a PII check tool. Maybe in my prompt I will say, like, please

check PII before you set anything out, which makes it more likely. I mean to do it, but put in your prompt whatever you like this thing to do. But in the end, like if you give it like a free set of tools, but you have specific combinations of those that you want to enforce, that's exactly the type of thing we can do.

Bart (38:51)
think that is also why these type of solutions are very relevant in a sense that the example that you gave for the healthcare insurer, I

think it's very recognizable, like where you used to have these support agents or bots or whatever you want to

Maurits Kaptein (39:05)
Yeah.

Bart (39:05)
call them. What we used to do is if like this tree-like approach, like the customer says,

Maurits Kaptein (39:10)
Yeah.

Bart (39:10)
we want to do this, I want to look into my account, I want to

Maurits Kaptein (39:13)
Yeah.

Bart (39:13)
look into my balances. based on that, you end up somewhere in the tree, you branch out.

Maurits Kaptein (39:18)
Yep.

Bart (39:19)
But like

Maurits Kaptein (39:19)
Yeah.

Bart (39:20)
all of those different parts you have as a developer of that support, you have a lot of control over and you know what happens, you know what the risks are for the independent parts. From the moments where we are now, like with LMS becoming way more performant and for lack of a better word, more autonomous. Like you don't

Maurits Kaptein (39:33)
Yeah.

Bart (39:34)
need the whole tree, you just need a very good prompt access to all the tools and the agent can basically do everything for you. But at the same time, as the person responsible for that agent, like it feels

Maurits Kaptein (39:43)
Yeah.

Bart (39:44)
much, much more risky because you have much less control over.

Maurits Kaptein (39:47)
Yeah

Bart (39:48)
What is happening, right?

Maurits Kaptein (39:48)
yeah I

I I'm assuming you played with these things like personally and then like yeah now I'm giving access to my credit card. Okay, let's let's see

Bart (39:54)
Exactly, exactly. Yeah.

Maurits Kaptein (39:56)
what where this is going. no but I think the the first thing you were saying, like we we work with a large bank in the Netherlands as well and and one of their effectively like the the way when we came in, like when we first started our conversations with them.

they already had, like I think every bank has a bunch of like automated workflows basically.

Bart (40:17)
Hmm.

Maurits Kaptein (40:19)
you you could like you know for whatever apply for a new credit card online and there was a bunch of steps that you would go through and that was just an automated workflow, it was fully scripted, nothing AI, you know, to it, just like that that there was an automated process that they had fed it and worked and whatever. And

Really, their first approach, which is kind of in some ways, even I would say sensible and low-hanging fruit, was just putting an LLM in front of those existing processes to kind

Bart (40:47)
Yeah.

Maurits Kaptein (40:47)
of like have a conversation with you about your intent, and then just route you to one of those

processes, which kind of like combined having you not needing to search through weird things and kind of having you know somewhat of a conversational experience, but you were still in this controlled flow, and every step step of it was controlled.

I think this is kind of like. this is the first steps that people are building. They're not, I would say they're not really building agents, but it's all like what's in the name, basically. But they're building like workflows that have LLMs, and maybe the LLMs like switch between you know two branches of a tree, basically. but I think right now, and this is obviously what these coding agents like Cloud Code, the Cursor, and these kind of things have been doing for quite some time, it's like

a very different open-ended approach. It's just exactly what you said. It's a powerful LLM with a set of tools and whatever task that decides like the LLM decides the full you know walkthrough basically.

Bart (41:46)
Yeah.

if we look more at the commercial side of this, like,

Maurits Kaptein (41:49)
Yeah.

Bart (41:50)
I'm also wondering, like, who are your typical customers in a sense that, what I notice when I talk to larger enterprises, they are typically,

Maurits Kaptein (41:58)
Yeah.

Bart (41:59)
like, we're very much in a bubble, right? Like, but the typical

Maurits Kaptein (42:01)
Yeah.

Bart (42:01)
large organization, they're like, feeling sometimes two years behind what is the state of the art. Right. And, and, and I can imagine that it's

sometimes more difficult to sell these type of solutions to large enterprises versus to younger companies that are actually building models and selling those to the building agents and selling those to the enterprises like companies like like in Belgium, for

Maurits Kaptein (42:23)
Yeah.

Bart (42:23)
example, you have Ravirkel and they're basically their end product

Maurits Kaptein (42:25)
Yeah.

Bart (42:26)
is an agent that they sell to enterprises, right? And I think that these can also be very interesting

Maurits Kaptein (42:31)
Yeah.

Bart (42:31)
customers for a product like Kyvvu. Like what is your typical

customer audience?

Maurits Kaptein (42:37)
Well, okay. I can do what I would like pitch to a VC if I I pretend to

Bart (42:41)
Mm-hmm.

Maurits Kaptein (42:41)
be fully confident, but may maybe kinda more interesting it's just the real story as to where we are and how

Bart (42:47)
Yeah, yeah.

Maurits Kaptein (42:47)
how it evolved. and it and it obviously that it's also just evolving very quickly. our initial

approach I would say was to talk to relatively large regulated enterprises.

Bart (43:02)
Okay.

Maurits Kaptein (43:02)
Why why regulated? Because like they they cared about like the traces, they

Bart (43:06)
Yeah.

Maurits Kaptein (43:07)
cared about the behavior. Like a lot of like smaller companies maybe well at least this was kind of one of our assumptions at the time didn't in some ways didn't care so much basically. Whereas like the these large enterprises often both had a combination of

Now very high level, top-down. We need to do stuff with AI, we need to do stuff with agents, but obviously we need to be responsible. whatever DOLGA, the AI Act, all of these things we need to comply, blah blah. Okay, so so

Bart (43:35)
Yeah, it's recognizable.

Maurits Kaptein (43:37)
that was kind of like the the story there, and initially we got

into some of those predominantly kind of in AI transformation, data strategy kind of roles, people that were responsible for like you know transforming the company using

Bart (43:54)
Hmm.

Maurits Kaptein (43:55)
using AI, using these tools. but that very quickly led to

The question being much larger than just our product, basically, because people are then going,

Bart (44:03)
Yeah, yeah.

Maurits Kaptein (44:05)
how do we build these agents? What's our strategy with these agents? What policies do we write for the use agents, which then is relevant to our product product, but not not necessarily core of what we would want to do. So we started working with partners.

In Belgium, for example, Datashift is one of our partners, whom you

Bart (44:22)
Okay.

Maurits Kaptein (44:22)
you guys might know, who basically are able to take our product, integrate it into the agents that they build for customers, and even

go as far, and and this is maybe a bit more technical detail than necessary, but obviously like endowing the harness with meaningful semantics. Um, so is this internal? Like it will how is this data classified? And writing the policies, like they can they can basically do that whole integration for for customers. in the Netherlands we also sort of have two partners that are doing the same and and we we kind of had this motion.

Where initially we interact with some of these larger end customers because you have to like generate some demand but quite quickly go to partners. And then if a partner works, you can see a partner like scaling up and offering it to their other customers as well. to be honest, that motion I I would say if you would ask me in my previous company, within the n like the nine months, not even that we're like the working.

It's amazingly fast. It's working very well. And we've gone through like procurement of some of these larger

Bart (45:25)
Okay.

Maurits Kaptein (45:25)
customers, because obviously that whole thing takes a long time and whatever.

Bart (45:29)
Definitely,

Maurits Kaptein (45:31)
But here's my my butt which actually ties to also what you're saying and it ties to just the world changing. What we see right now is

Yeah, well one way of putting it, but that's maybe a bit negative, is I would say if we continue this way, then next year we'll be in let's say a hundred, maybe a thousand, but kinda order of magnitudes agents at large enterprises.

Bart (45:54)
Mm-hmm.

Maurits Kaptein (45:55)
Which is great, that's that's nice, you potentially make some money and whatever. But by the time next year there will be like a billion agents. And then like,

Bart (46:05)
Yeah.

Maurits Kaptein (46:05)
how this we we did something wrong if we're only like in in like a thousand of those. Like I wanna be in all of those, or really like we should be the standard that governs all of those. So what we're really now trying to do, like and one of the things that's actually gonna be an active push in the next you know, two to three months is the group

That

you're just alerting to, like the people, the smaller companies, but that they're basically building agents, or even wider. I think everyone, well, not everyone, but a lot of people that offer it software products are now

Bart (46:37)
Mm-hmm.

Maurits Kaptein (46:38)
putting like agentic type you know interactions into their products basically.

Bart (46:41)
Yeah, definitely.

Maurits Kaptein (46:42)
so that that's definitely a push that we are starting to do. we'll see how that goes. Um

And this is kind of I think where for now we're we're gonna be focusing kind of the developer side, people building these things, how do you build this into there? But then there's an even larger discussion that we have. I don't know if you want to kind of like go into some of these details of st of startup life and how do you build a company,

Bart (47:08)
Yeah, it would be interesting because I also

Raphaël (47:09)
course.

Bart (47:10)
find it a very interesting reflection where you say we're growing quite well, but at the same time, the market is growing way quicker. So actually something needs to change.

Maurits Kaptein (47:19)
So we're not fast enough. Yeah, yeah, that's just

this is that well, that's that that was the first observation. The other one, which is now which is also quite interesting, I think, as a way of like abstractly thinking about this, like, we secure agents. That's like the high-level story. We've already covered some of the nuances of this. Like, but but hey, that's the high-level story. This is just how I have it in mind. So now I'm gonna like draw something in in the sky. Like, we secure agents is a high-level story. technology stack is kind of

It's what we call our engine, it's the thing that actually checks these policies. And then we have a platform for assigning policies and that's like the core of our stack is well, I don't want to say fixed as in like in in a year from now we'll be the same, but like there's a clear core that just given the behaviors, check these policies, we standardize behaviors across harnesses, blah blah. So then you have these two endpoints.

And what we're the way I'm thinking about is like, oh, now we're really focusing on people that develop their own agents. And starting with these large customers and and this partner motion, maybe reaching out to some of them directly, like smaller companies that are building agents. I think that there there's, you know, a large win there. By the way, then you also have to think about like how do you actually

what's the business model, which might be quite different from from even between those two groups. But on the other side of this like circle that I have in mind, so this is want secure agents, we have the same tech stack, agents we built ourselves, but there's also third-party agents, agents that you buy.

basically agents that

Bart (48:50)
Mm-hmm.

Maurits Kaptein (48:50)
you use. This might even be like your your co-pilot instance basically or your cloud code instance or whatever like well what meta muse and and these kind of like the

Bart (48:59)
Yeah.

Maurits Kaptein (49:00)
agents that you buy potentially all the way up to consumer but that's not something we're entertaining right now. But there is like in most of the conversations that we have with these big companies there's

Often a CISO or somebody from IT, but often it's a CISO-type role who's worried about all these third-party agents that their workforce is

Bart (49:22)
Sure.

Maurits Kaptein (49:22)
using, basically. And that's there's a there's a place there where potentially, like with the same core and with the same story, you would also be selling.

to to that kind of segment of the world. but I do think like I'm I'm a bit reluctant. We we we're having this discussions. I'm a bit reluctant because technically it's a quite a different stack. So in in a

Bart (49:43)
Mm-hmm.

Maurits Kaptein (49:43)
sense like on this patch you build for a great like developer experience, which is I think what we're trying to do now. And on this side you have to like double click on Windows or something. I don't like it's a different it's it it like it might still be the same core and the same story but it's a kind of it's quite a

different path in how you develop and and what you

Bart (50:03)
Yeah.

Maurits Kaptein (50:04)
test against and that kind of stuff. So right now I think we're we're like we're sticking with this like the developer side of things but likely like expanding not only through like partners these large regulated enterprises but also exactly the well the the group you were just mentioning like the people that are like slightly smaller but building agents or building agentic skills into their existing platforms.

Bart (50:28)
And

do you also see Kyvvu going, because you're now in this business of policy enforcement,

Maurits Kaptein (50:32)
Yeah. Yeah.

Bart (50:34)
like, do you also see it like capturing this more guardrail aspect or the sandbox aspect? Or do you see that really as separate parts? Because I can imagine, like,

as a customer, if I want to build a solution, it's annoying if I have to go for every little part because I need all the parts. I need a different

Maurits Kaptein (50:51)
Yeah.

Bart (50:52)
vendor or a different supplier, right?

Maurits Kaptein (50:53)
Fair question. Fair question our VCs are a asking us as well. It's like, why don't you become the like

Well it it's tricky 'cause I had like the I wanted to phrase it in a way that you didn't ask it. But I wanted to phrase it as like why don't you b to become like the the everything platform for agents

Bart (51:09)
Yeah.

Maurits Kaptein (51:10)
basically? I th you you you you ask it more nuance than this. But but I think like there's a few people obviously that are trying to build these or promote these platforms that like build your agent here and then we'll build it secure or like we're an agent gateway and by being a gateway we can do all kinds of things 'cause we can sniff the

wire and look at and and here's your solution to everything. to be honest, with the larger corporations that we interact with, which obviously self-select in a way, uh

Bart (51:38)
Mm-hmm. Sure.

Maurits Kaptein (51:39)
that they are all

Very much off the position that they they already have logging tools. They already have

Bart (51:46)
Okay.

Maurits Kaptein (51:47)
in part, like you know, the many of them are thinking about like doing content checks that are very specific for the company and being value inside of their company, basically. They want something that runs on premise, which is what we do. Like the engine runs

Bart (52:00)
Sure.

Maurits Kaptein (52:01)
next to the agent, the policies are retrieved, but can also be locally, so you can have a footprint where no data leaves

Bart (52:07)
Mm-hmm.

Maurits Kaptein (52:08)
your systems, there's no network hub, like the whole architect.

Architecture

that we have and the specialization for some people is is the appeal, I think, to our product.

Bart (52:17)
Yeah, yeah.

Maurits Kaptein (52:18)
But I I get what you're saying, and I think like in the end, like we either diversify, like add these things, or or ensure that we basically have good collaborations, or like it, you know, obviously, an integration partner would put everything together for the for an end customer. but for now.

And let's see how long that lasts. I want to make sure that the thing we do that we do it well. There's still enough challenges in in in even the thing that we do. And I'm I'm really but this is also just personal taste. I I I hate like platforms that do everything. I I

like products that are like clean and do something well, basically.

Bart (52:56)
Yeah, yeah. So there's a good

Maurits Kaptein (52:58)
Yeah.

Bart (52:59)
argument as any, right? I don't understand what you mean because then for this specific aspect, you can also make sure that you have a very good developer experience and like you can really like double down on being very good at this, right?

Maurits Kaptein (53:09)
Yeah. Yeah.

there is a a thing that we are entertaining, which I think is quite cool and would like in some ways basically enforce the the the way we're doing this is like well but you you're saying why not also offer to people that are building their own agents that they sell.

I think

Bart (53:33)
Mm-hmm.

Maurits Kaptein (53:34)
yes, that's a that's a thing we we will actually start doing. however, should they make their policies or should the buyer make the policies? So now there's this interesting flip

Bart (53:44)
And raise a question.

Maurits Kaptein (53:46)
that you can do where you can say, look.

I am whatever the some big bank in the Netherlands. I'm gonna buy this like third-party agent, like a customer support agent built by some company that built, but I'm only gonna buy it if I can be sure that it listens to my rules. So not the rules that they set, but I set those rules because I set rules for all of my agents. Now you get kind of a weird split in in our product, but but I think it's one that's that's quite interesting. Where where you like the flip around is like, look.

I will only use agents inside my company that effectively like emit behaviors the way that they're checked by Kyvvu. otherwise we won't be buying them, which is obviously kind of an aspirational point for us to be at. But I I do think it's it's a it drives a different type of value, basically.

Bart (54:36)
It's interesting, yeah, because if you position it like that, that like you sell to companies that are selling agents, but that also means that it means for these companies that they actually, there's more value in it to them because they can sell a new feature to their customers, right? Like they can sell the feature of, can now define your own policies on these agents.

Maurits Kaptein (54:56)
Yeah.

And and now you can also talk about and this is what I sorry, I just slipped that in in one sentence about what what is your business model because now like but effectively right now

We sell through partners through these bigger companies. The end customer pays effectively a usage fee. Basically, every active agent that's being governed as long as it's running, there's a it's very small, but it's a fee for active agents, and and we we do this under the belief that there will be many active agents with these companies. But for

This other route, targeting people that are building these agents themselves, well, you're providing them value in a few ways. One is like, are you giving basically you're taking away some of the engineering that people will start to expect anyway? Because even like I mean, if if you run whatever, I happen to be on a bunch of anthropic stuff myself, but if you use cloud desktop

It asks you for permission, it asks you for rules. Like,

Bart (55:52)
Is it good?

Maurits Kaptein (55:53)
this is what people are starting to expect anyway. So anyone building agents would have to build this, which is something we would provide them out of the box. We would provide them with traces out of the box for the AIX, these kind of things. And indeed allow their customers to set their rules.

Bart (56:09)
Hmm.

Maurits Kaptein (56:09)
However, business model wise, now you can even imagine why would I charge these people? Why would I not actually like just take a cut from the end customers so they actually make money on integrating us into their agents, basically. And this

Bart (56:24)
Yeah.

Maurits Kaptein (56:25)
is that's a very different discussion, but it's like, but the moment you change that target group, now you have to also start thinking about what business model actually works for that specific group? So

Bart (56:36)
Yeah, indeed. if you change your target customer, if you become like a part of their product, like

Maurits Kaptein (56:42)
Yeah.

Bart (56:43)
it becomes interesting to think about a pricing that's always a win-win for both of these parties, right?

Maurits Kaptein (56:50)
Yeah, and then and then I and I think there's there in in fairness it works in all directions basically because

Bart (56:56)
Yeah.

Maurits Kaptein (56:56)
your end customer gets this capability basically that they have and

Bart (56:59)
Exactly.

Maurits Kaptein (57:00)
for for us, doing some revenue split is also a way to incentivize people to build us into their agents, which obviously is is of value for us. So no.

Raphaël (57:10)
Moret, you spoke about uh VC and the discussion that you had with them.

Maurits Kaptein (57:16)
Yeah.

Raphaël (57:17)
So you raised one million in last August

Maurits Kaptein (57:20)
Mm-hmm.

Raphaël (57:21)
in the Cedrons. It was led by Vault Aventures, a Belgian VC,

Maurits Kaptein (57:24)
Falta ventures, yeah,

Raphaël (57:27)
by the way, and also a public agency in the Netherlands.

Maurits Kaptein (57:32)
Yeah, the the the yeah.

Raphaël (57:34)
You mentioned the discussion with these VCs. How is the appetite now of these funds in terms of AI security? Because this is a hot topic today.

Maurits Kaptein (57:49)
Yeah, oh okay, that's it. Yeah, so so Fort Aventures by the way is from Belgium and they're very active in the Netherlands as well. Brom is Brabant's Wikipedia who funded us, and we we also have a bit of money from the Rabobank. so that's kind of the group put together. but yeah, I would say look, if you take a very high level, this is partly also my own opinion by the way, but

So it comes forward in some of these discussions, it's like, this whole AI game is moving. People have like the the appetite to just invest into any anyone that puts AI onto their product. That's that's I think gone. Um

Bart (58:27)
Mm-hmm.

Maurits Kaptein (58:27)
but there's a few things like obviously if you're in a specific market, if you understand like a specific you know workflow very well, I think that's still that I mean that's a place where where where people can make money and would invest. I think

investment in us and we we are part of a larger category of i guess emerging problems so and really i think there's there's three very clear emerging problems with like these agents that are coming up one is how do you check the quality of what your agent is doing so so the whole benchmarking how do you make sure this thing is actually like the output is good

I think there's an appetite to to know invest in in things that actually solve that problem because I don't think it's properly solved. you do see it popping up everywhere. The same for well security. I think that's that's where we are. And I think right now there's definitely people willing to invest in agent security.

I do feel it's gonna be a bit of a a race in some ways. This also ties back to why I think that now what we're doing now is potentially too slow. because I think this security, especially the level that we're doing it, in my view, this is gonna be common p plumbing, basically. And and it's gonna materialize in

A year, maybe two, but not much longer. And then there's a lot of incentives for this to be very standardized. Because it's there's a lot of incentives for every agent to emit their events in a standardized way, is all of these things. So so you know, we we have to move fast. But I I think generically there is an interest in in funding security for agents. Um and then the last out of these three problems that I also see that that that people do have an interest for.

But it's not solved at all, I would say. is accountability of what these agents do. So

Bart (1:00:15)
Yeah.

Maurits Kaptein (1:00:16)
who's responsible? Are we gonna insure this? Or how like how how are we gonna work with these

Bart (1:00:19)
Good question.

Maurits Kaptein (1:00:21)
things basically? that that's another question that's very open. Um I think in general VCs are willing to invest in in in good stories or in good products, but they also have a hard time because they see the world changing and they see how fast this is moving and like the normal time frame would be whatever.

you know five to seven

Raphaël (1:00:38)
Yep. True.

Maurits Kaptein (1:00:40)
years it's very hard if you it's like I've had these discussions when we were raising this round with VCs people were saying like where do you want to be in five years I was like I hope to be still alive but I that product wise no clue I don't know

Bart (1:00:52)
Yeah.

Maurits Kaptein (1:00:53)
what this world is gonna look like so this is this is a problem for entrepreneurs but also for VCs I would say yeah

Raphaël (1:00:59)
And I think that VC

Bart (1:00:59)
Yeah.

Raphaël (1:01:00)
also wants to invest in people in this context, I think that both Volt Aventures and BOM, which is the Dutch

Maurits Kaptein (1:01:07)
Yeah.

Raphaël (1:01:08)
public agency, were also

Maurits Kaptein (1:01:10)
Yeah.

Raphaël (1:01:10)
investors in your previous company, if I'm right.

Maurits Kaptein (1:01:13)
So that

so that helps. So we didn't we I I want to say we didn't put a lot

Raphaël (1:01:16)
That helps, in fact.

Maurits Kaptein (1:01:19)
of effort in fundraising, which then sound sounds really weak, but but yeah, we we did a very small round with a few people and like Jeroen, like our CEO knows Folta very well again because he was on the board through Folta, because like one of the people that runs Folta was an early investor in his previous company, not through Folta through

Bart (1:01:38)
There you

Maurits Kaptein (1:01:40)
another fund. but yeah, in the end I think

I d

I do think this is a large part that drives. It's like, you've worked with these people before, you've seen them be successful before. Yeah, obviously that's a big help at this stage. Yeah. Yeah.

Raphaël (1:01:52)
I don't think that it sounds weak. In finance,

we sometimes forget that investing in people we believe in is sometimes more efficient than analyzing the market

Maurits Kaptein (1:02:04)
Yeah.

Raphaël (1:02:04)
and going through heavy due diligence to choose

Maurits Kaptein (1:02:07)
Yeah.

Raphaël (1:02:08)
your investment. Sometimes you just have to follow people.

Maurits Kaptein (1:02:11)
Yeah, I d I think this is kind of interesting, like uh Sandra Funk from Folta, we've had many discussions also in the previous company. I think one of the things

My takeaway from how he speaks about this, especially like they do a lot of very early stage like tech investments. and to him, like, it's the people for a large part, and it's the market. But the market is something that like his the identification of the market size and the opportunity and that kind of stuff effectively comes from him, like not necessarily from the company. he just he knows like agent security.

or AI security, that's a that's like a space that I wanna invest in. And then if there's the right people that are working in that space then then that's a match. But the like the the numbers effectively like or the modeling that a company does at that early stage, it's like yeah.

Bart (1:03:04)
I think this space

is very interesting in the sense that like AI security, if you talk to any CISO, any CIO of a large organization, it's top of mind. But at the same

Maurits Kaptein (1:03:13)
Yeah.

Bart (1:03:14)
time, there are no clear winners yet. There are no de facto off the shelf things that you can do to you can

Maurits Kaptein (1:03:16)
No. Yeah.

Bart (1:03:19)
buy it to be okay, right? Like with if you want to build tomorrow a

Maurits Kaptein (1:03:23)
Yeah.

Bart (1:03:24)
data lake and you go to Databricks, there's...

Maurits Kaptein (1:03:26)
Yeah.

Bart (1:03:27)
That's the right choice, You can do like five right choices and you're always gonna be okay. But here there

Maurits Kaptein (1:03:29)
Yeah,

Bart (1:03:33)
is not really, there is no best practice, there is no off the shelf, this is the way you do it.

Maurits Kaptein (1:03:37)
Oh, it's it's true.

I mean that's that that's the interesting thing. This is like one of the things that Jagoon also like the the way he talks about just our commercial process, which I think is kind of interesting. It's like, yeah, right now what we're selling is like we're selling this like generic idea. Like people care about this idea, but now you have a lot of like f fake competition and that's what it what he would call, like other people that also sell to this idea but actually do something very different.

And he's like and it's very different from selling a phone, for example. By the time like a product b has become like a known category,

Bart (1:04:13)
Mm-hmm.

Maurits Kaptein (1:04:13)
now your selling is like it's competing based on features and these kind of things. But right now we're kind of like

pre-feature stage basically in in in many ways. Like nobody's

Bart (1:04:22)
Yeah, yeah. That's interesting.

Maurits Kaptein (1:04:24)
asking us how quickly do you evaluate these policies, or like just properties of the of the product in some sense that we do care about are not driving the sales conversation yet because the the category just doesn't exist yet or like the they haven't materialized yet basically.

Raphaël (1:04:42)
um

There was a question that I wanted to ask you. It's maybe

more general points because I'm sure that not all of our listeners are pro developers and a lot of them use just codecodes

Maurits Kaptein (1:04:59)
Yeah.

Raphaël (1:05:00)
or other agents because they want to test or to save time at work or

Maurits Kaptein (1:05:04)
Yeah.

Raphaël (1:05:05)
in their personal life to send email, wire money, trip, I don't know.

any kind of things, they

Maurits Kaptein (1:05:12)
Yeah.

Raphaël (1:05:12)
are maybe not aware of all the security checks that they can do. So if you had to give them a few pieces of advice before they let an agent on their computer, what would

Maurits Kaptein (1:05:27)
Yeah.

Raphaël (1:05:28)
it be? And maybe this is a question outside of the Kyvvu context.

Maurits Kaptein (1:05:32)
Yeah. It's it's an it's actually an interesting question. 'cause it I should be able to just fair like give a very crystal clear answer. And and I and I don't think I can because it's like 'cause it's moving so fast, like I think

Half a year ago we're doing these first agents and and yeah I was also sandboxing them like on a different machine and on like different accounts and and this kind of stuff basically. Right now, if I look just for myself, like the amount of access that I give to to some of these tools. we we do like dog food our own products so so we we don't have like our own products running on on some of the tools that we use.

Yeah. I would say at the very least like read what it's asking you for, if it's asking you to allow something basically. and yeah, but this is I think this is poor advice, but I but I I do think like some of the larger vendors, like the larger labs and the large they they are for their individual like

Cost consumer-oriented products are obviously also trying to make them safe. So you know this is like don't play around with too too many like a developer built this have fun kind of tools. If if those are the ones you play with, then make sure you really

Bart (1:06:54)
Yeah.

Maurits Kaptein (1:06:54)
sandbox and do all of that. but it's it's changing very quickly.

Yeah, I I I would right now I don't give for example like I don't give any like access to payment systems and these kind of things. I'm very reluctant with just blanket like browser access basically. So so that's something that I I always do shield off.

But yeah. Again, not the not the best crystal clear answer because the the world is like every week there's new yeah.

Raphaël (1:07:20)
We'll take that. We'll take that.

Maurits Kaptein (1:07:22)
Well, I think the like one of the things even like cloud desktop, I don't know if you guys are using it and and I'm sure like Entropic and then Google and whatever have competing product. But about a week ago or something, two weeks ago, they they changed quite a bit in how they do their

Yeah, there's a system for for you allowing things. They started in some ways like the the the very noticeable thing was like they started asking for way more permissions basically. And

Bart (1:07:48)
Hmm, yeah.

Maurits Kaptein (1:07:49)
it's it's not as easy to give blankets permission for two things anymore. so you see I think like generically you just see d a a tying up of things because yeah,

Bart (1:07:58)
Hmm.

Maurits Kaptein (1:07:58)
we we're starting to see the errors, so we wanna wanna prevent them.

Bart (1:08:02)
but I think personal experience on this. Like

Maurits Kaptein (1:08:04)
Yeah.

Bart (1:08:05)
if I look at myself, like the more you start trusting it, the more nothing goes wrong. Like the more access

Maurits Kaptein (1:08:11)
Yeah.

Bart (1:08:11)
you get, right? Like also if you've done something two times, you say allow once, allow once, like the fourth time you're gonna say always allow. Like you get these type of, I noticed myself and I'm not sure if it's smart that I give.

a lot of access to a lot of things. even browser usage, I use it a lot to automate stuff. The other day, there

Maurits Kaptein (1:08:31)
Yeah

Bart (1:08:31)
was a startup that asked me to look at their product and I just pointed codecs to it, to my browser to basically go to basically speed run to

Maurits Kaptein (1:08:39)
Yeah.

Bart (1:08:39)
the replication and test everything. Like, and it makes you super

Maurits Kaptein (1:08:41)
Yeah.

Bart (1:08:42)
powerful, but at the same time, you need to give a lot of access to these agents, right? It's a difficult balance.

Maurits Kaptein (1:08:47)
Yeah.

yeah, and it's also a balance of what like what are the costs of some of these things. Like you know it's so some things are not

Not too bad if they go wrong, basically, and and some things are. I think well one thing that we're doing, for example, internally in the company is that we've by now moved virtually everything to Git to basic like to version controlled systems, basically. Because we allow like our agents to build sales decks and then and everything. but then at least we wanna be able to properly roll back and have for like then it's a it's a very different way of like addressing some of the problems.

Bart (1:09:23)
If a trace, yeah, true.

Maurits Kaptein (1:09:26)
But it's also just like making making potential making potential errors less harmful, basically.

Bart (1:09:28)
A way of keeping control to some extent. Yeah.

Raphaël (1:09:35)
think that's a good way to conclude the podcast. um Thank you very much, Moritz, for your time and your answers. Maybe if people are interested to know more about Kyvvu and its

Maurits Kaptein (1:09:48)
Yep.

Raphaël (1:09:48)
future development and maybe about yourself, how can they follow you and maybe reach out to you?

Maurits Kaptein (1:09:54)
at kyvvu.com. This is like the the easiest place to find us. We we kind of actively try to even blog about some of the things that we just see happening.

Raphaël (1:10:03)
Very interesting blog post, by the way.

Maurits Kaptein (1:10:05)
Well, it's just also stuff we see happening in the space basically. so we're tr we're trying to just also have an opinion there. you can find me personally on LinkedIn very easily and I'm I'm I'm quite

I didn't end up active there, I don't know what that means, but if people reach out to me there in in DMs, that that that's stuff that I see. So I'm happy to reach out that way.

Bart (1:10:26)
And I think you also have very intuitive docs on kyvvu.com. ah for people that want to test it out themselves, it's

Maurits Kaptein (1:10:32)
Trying to

Bart (1:10:33)
quite easy to get started, right?

Maurits Kaptein (1:10:35)
It is actually like uh so we we do it this is something that we haven't really pushed at all, but we we just have like a full run through like if you're an individual developer, you can actually use Kyvvu for free for pretty much like an active agent yeah all the time basically. You can even run it fully locally and our docs, yeah. I I think we have fairly simple examples

Bart (1:10:58)
Hmm.

Maurits Kaptein (1:10:58)
to get started. we also have I think fairly good yeah the examples I could

enough to get your agent to get started with us as well. So that's the that's the other side

Bart (1:11:05)
Yeah.

Maurits Kaptein (1:11:07)
of things. so if you want to try Kyvvu you can just point your agent at it. So

Bart (1:11:11)
Thanks a lot, Maurits.

Raphaël (1:11:11)
Thank you Moritz.

Bart (1:11:12)
was very interesting. um Thanks a lot, everybody, for

listening. And we'll hear you all next time.

Maurits Kaptein (1:11:19)
Cool, thank you.