Allie Howe Hey everyone, welcome to this episode of the Insecure Agents Podcast. I'm your host Allie Howe and today we have Ivan Burazin who is the CEO and Co founder of Daytona. Ivan, Would you like to introduce yourself?

Ivan Burazin Sure. Thank you so much, Allie, for having me. Yeah, CEO, Co founder of Daytona. Daytona is a infrastructure company. Purpose built for AI. Agents, meaning we essentially give AI agents computers. To do their tasks. So basically the way you can think about. It is. Our number one consumer is the agent, an AI, like essentially an AI agent, less so a human. And so if we equate an AI agent to a digital version of, you know, a knowledge. Worker we A. Daytona sandbox is what? We call it is. The equivalent to kind of like a computer, a, you know, MacBook. Or a laptop or. Whatever it may be.

So if. An agent has. To do a task like. Code execution cloning GitHub. Repository it has To use a browser use a. Computer use we enable it to very simply spin up these computers to be able to execute on those tasks and so the most. Visual way for people. To think about is if they've ever. Used like ChatGPT and give it like an Excel file or some complex mathematical equation. It'll go like thinking and it'll spin up this type of machine to execute Python code against that. Or if they've used like the agent mode, like open up a browser and scheme across the web. That is also a similar use case.

Now OpenAI is not a customer, just to be very clear on this. But that is like the use cases that. People use Daytona for.

Allie Howe Yeah. I've heard people call you the sandbox company. I heard on a podcast that you did on the Infrapod with Ian and Tim, you said that Daytona is kind of like the PC or the Mac for the agent instead of maybe just like a sandbox. Can you elaborate on that?

Ivan Burazin Yeah, because of the name the the word sandbox is the name of the primitive. That we offer so. If you think of, you know, an AWS like there's an EC. 2 which is the. VM that they offer you. The sandbox is the type of VM that we offer for agents and now the sandbox is the name that the market has consolidate on that they will call it a sandbox. Although the sandbox I feel is quite a limiting name and also. Feels about like. Testing and not actually in production type things so. I do try to. Explain that it's more of a full on computer for an agent than an actual than something something that you think of as a sandbox and more specifically we think of.

It as a like. Composable computer for the agent. What I mean by composable computer is that your agent can, depending on the task on the fly, define this computer like oh I need this much CPU, I need this much RAM, I need this much disk, or I might need a GPU, or I might need a windows machine, or I need a Mac or a Linux, right? So it the agent can on the fly define these things. Sort of like if. You can imagine yourself like in the old times like in a PC shop like putting a PC. Together with the. Pieces installing it. That is what we do.

For the agent, but. Very, very fast. Like in sub 100 milliseconds, the agent gets this computer up and running.

Allie Howe Wow, that's insane. Is that kind of like the secret sauce thing? So I can imagine like I myself have tried to like host like local models and AWS by spitting on my own like EC2 machine and bring it like, I don't know, a CPU or whatever it needs. And it takes just forever for that to happen. It slows me down a lot.

Ivan Burazin Yeah. So quite similarly to that. So your agent, the way I think about agents is agent essentially, you know, it's a piece of software that's connected to a model, whatever model from Open AI or Anthropic or anyone else. And then it has the ability to call on tools and then the Daytona machine can be a tool for it to do something. But also what's? Coming more and more. The use case is. Because we have things like Claude code and these other CLI agents that are made to run on your machine, then people have now started using these sandboxes. To actually throw these. Agents inside of it.

So now they'll spin up a Daytona sandbox with Claude code inside and then it'll be sort of quote UN quote free but locked down to do what it wants. Like it won't break your laptop. You can spin up multiple of them. You can set some sort of limitations on like firewalls, what can it access and things like that.

Allie Howe So it sounds like the primary reasons that an agent would need a sandbox are speed, statefulness, and security. I was also wondering like how Daytona helps with security for agents by giving them the sort of those firewalls or rule sets baked in that make the agent do only things it's supposed to. And like does it help with adding hard boundaries?

Ivan Burazin Sure, I mean the number one use case why an agent would first need a computer. If it is. Depending on the. Task it has like. If the agent is not running inside the sandbox, then it will only call the sandbox if it needs to do that. If you ask an agent, you know what is the capital city, whatever, it doesn't need a sandbox for that. But if it's like. Oh, here's a bunch of XL files or CSV. Files now go analyze this then it does need it and it'll spin it up and that's if the agent is outside of it. If the agent is inside of it, so a cloud code running inside of it, then the security part becomes actually quite important, right?

Because now you have a agent inside of a computer that has visibility to everything. So if you have like secrets and tokens inside of that, the agent can see it. The other thing is. So you have to. Like solve secrets management for the agent or do some sort of like proxying of these API. Keys, so it doesn't. Go off and do, you know, malicious things? And also something that people have asked quite a lot about, which we've launched quite recently, is, you know, firewalls inside of these machines. And the firewalls are both like incoming and outgoing. So it's like, where can the agent go on the Internet?

Like you can define that and also who can, you know, see what is being published inside of these sandboxes? So, yeah, that's definitely one thing that has become security has become more and more important, especially now that agents are running inside of these machines and can actually, do, you know, whatever they want.

Allie Howe Do you ever think of Daytona as a security company?

Ivan Burazin We're getting flagged like that more and more, although I've never. Thought of it. As a security company, we mostly think of Daytona as a performance enhancer sort of company. So basically the way we think about it and what we've seen and working on more detailed benchmarks around this is that your agent is probably. At least double. If not orders of magnitude more performant using a Daytona sandbox versus, you know, a Docker container, a Kubernetes spot, or whatever. Just because of how fast it works the. Statefulness, the tools that we have in all these things basically save time and tokens and context window for a agents enabling them to be on average more performing than anything else and so.

The way we've always looked at it was like. Sort of in that sort of performance vector. But because of the. You know. Aspects that. Agents can do, and the fear of what they can do. Security is a very, very, very large. Portion of our product. I'd say like at least a third of our product is like focused. On security for sure.

Allie Howe I find in my experience, that's almost one of the best ways to sell security, where like security is kind of a secondary outcome, but you're selling that performance, you're that reliability piece. When I think about trustworthy AI and what that is, it's not just AI that's secure and safe, but it's also AI that's performant and reliable. And if you can deliver all of that, then I feel like you've got like a really powerful go to market strategy.

Ivan Burazin Exactly because ideally you're trying to solve a problem for your user and or buyer. So it's like your user and or buyer wants their agent to do X or Y. And if your agent has a higher probability of doing that vis-à-vis, then that is a great thing. And then at your as you mentioned on top of that, if it's. Also like secure or. At least more secure than that is like an extra win. So yeah, I agree with you 100%.

Allie Howe When you spin up a Daytona sandbox, what sort of like comes out-of-the-box for security? Others firewalls, others rule sets? Are they configurable? Are things already sort of set up for you? What does that look like?

Ivan Burazin Yeah. So there's a lot of things that are default. And so when we spin up a Daytona sandbox, the thing that we pride ourselves on is. What we call the. Agent Toolbox, meaning if you think of a computer when you buy and you want a laptop like no matter what it is like windows, Mac doesn't matter when you open it fresh, it still has apps in. It right it sounds like a file. Explorer, we don't think about this. An app like a terminal, that's an app, right? Enables you to, you know, get things done on these machines. And in that same vein, when an agent spins up the default data on a sandbox, it has these things like a headless terminal, a headless file.

Explorer headless LSP like headless git. Client and so these things come essentially out-of-the-box to make your agent work better. Now as time has progressed, we've added new things and let's say for example. The git client itself already obfuscates. You're, you know token so like the agent can see them which is great. The firewall, basically. Everything is open default, but you can define. Exactly what? You want at creation time. So essentially the human, we're going to call it owner, creator of application and or agent can add the constraints that when a sandbox is created, it always enforces these rules, right? And then your agent is sort of locked down by that.

So the firewall is in there automatically. But the firewall rules are up to the. Integrator owner of the system to define.

Allie Howe Do you feel like most of your customers today are customers that are calling out to a sandbox for specific functions like analyzing data or browsing the web? Or do most customers like host their agent just all the time within that Daytona sandbox?

Ivan Burazin The latter is newer, so historically the vast majority of customers the agent is running somewhere else. And if you think of like any of these vibe coding tools. Just to be very visual if you. Log into a vibe coding tool you have on the left hand side, probably like the chat interface, which is how you interface with the agent. Then there's like this like iframe type thing where like it shows you the code that's ringwrited or previews the outcome of the project and sort of like that iframe sort of is what the Daytona sandboxes and then the outer part is what the agent is. It's probably running on, you know, AWS or or something like that.

Historically, almost all the customers are using it in that form. But with this. Proliferation of these new, as we mentioned, like CLI coding tools. I feel that people are like, oh. Why go out to build all these things? I can just like throw one of these agents inside of some sort of secure sandbox, send in a command and I can spin up as many as I need till I get a successful outcome. So it's like it's not one agent. It's like, oh, I can throw 50 of them, 100 of it doesn't matter, right? Just like throw until you get a successful, you know, outcome that you want.

So this seems to be a new sort of shift in the paradigm. And it's something that when you're building net new products or products that are at the beginning of some market trend, it's really important to, you know, have your ear to the ground and start releasing. Features because firewalls. And you know, secrets management and things like this weren't really interesting a few months ago because the vast majority of agents were like outside and it was like a different problem set. You didn't have to solve that. Now that it's more common to put it inside or the newer products are more common to put inside, now these are the new problems that we have to.

Solve on security right? So to answer your question, it seems that directionally it's going to. The. Ladder where agents will be inside the sandboxes. But the vast majority. Of customers are still. The former.

Allie Howe Interesting. Yeah, that's awesome. I saw that you are growing like crazy on Twitter. I saw your post where I think you almost had like 2 million sandboxes a day. What do you attribute to that recent spike?

Ivan Burazin It's interesting to. See how these spikes? Go and it's hard to even imagine how they keep going the way it. Works or looks or feels for us is you start off. And then you have like some sort of linear growth like day over day, you know, it's slightly up, slightly down, but generally trending upwards. And that is, you know, you have a a few, you know, days or. Weeks that are like that. Then you have this, you know, new customer or existing customer that did a launch or whatever, and then there's like a. A spike that's like. 5X10X larger than whatever you were doing on average before, right?

And then it just stays there and then it's just literally growing again because now that product is launched and they have their organic growth, right? And then you know, a week later, 2 weeks later, you have a new customer, a new product they launched. It's like 5X more or 10X more. And then they just keep adding up like that. And it's like funny to see how like when you saw. That tweet like all our other like growth crafts you they're basically. A flat line. Like you don't even see them anymore, but when they were happening, they're like, oh, this is so huge. It's so big.

And now they're like. Because you keep like. Adding an order of magnitude more of growth and it's kind of on one side it's exciting, the other side it's like. Scary. It's like. Where does this end? Right, because we've been live for six months. So it's like still very, very new and this keeps growing at this trajectory, the numbers become just for. Lack of a better word, Stupid. It's just. Like insanely big and the other like the the other. Part of that. As and I and someone, one of our customers actually. Tweeted It's like we felt this. And this because that's not. Good, but we've solved.

That since where it's like every time something grows in order of magnitude, especially when you're new, hopefully it never happens again. But there's some things that become shaky and, or break and, or unexpected, right? Because linear growth is sort of easy to predict what is going to happen. But exponential growth is really really. Hard. So there's a. 2 ends of the spectrum. But to answer your question, the growth is attributed to new products and or new customers. Infra companies are. Basically very power law. Driven, so you'll have like a small set of customers that will drive the large portion of your growth and as we keep adding these.

Large. Companies, these things will continue to scale and if we look. At how fast? These consumer AI companies are growing as they start building on us or using US new ones around, they just grow faster and just like have this large impact on our, you know, creation metrics.

Allie Howe Interesting. Yeah. I was wondering like where the market is in terms of like understanding and education around when they would need something like Daytona. Like I wonder if most people today like start out with some like overcomplicated K8 setup and then they run into like stateful challenges and speed challenges. Like do they have to go through the pain first? Are there or is the market just like, OK, we know we don't want the pain, where to go straight to Daytona?

Ivan Burazin I don't think the market knows that and again, like we're doing really and hence we're on this podcast. Hopefully some people find out about this. So that is great. We're very new, like there are competitors in the market, but the market in general is still very new. And I'd say that the number one quote UN quote competitor. For us is. Like, you know, Kubernetes. Cluster. Where people are like, oh, my agent needs to run code, let's spin up a Kubernetes and they have a pod and and go. And the biggest thing that I've found with talking to these customers that end up converting to us, the first thing is scale.

It's just like. At some scale, Kubernetes pods don't work really. Well, because. They're made for something. They're not made for agents to work for a really long time to solve them. They're made for AB deployments or made for a very different use case. And so that's the number one. It's like, oh shit, this is great. I'm paraphrasing, but this is breaking. Let's find a solution, right? That's the number one thing. And then when they find plus hopefully then there's like, oh. There's all these like. Tools and I can run forever and all these other things that make my agent even more performance like 1. I don't have to like suffer breaking and managing this Kubernetes cluster, but also my agents are much, much better.

And so I think to your point. That is still very. True. And so we're doing as much as we can on the, you know, go to market awareness education side so that people actually know that we exist in what we do.

Allie Howe Is security part of that go to market strategy? Because isn't like Kubernetes really hard to like secure? A lot of teams like run into the challenges like that. I know Kubernetes is not secure, just like right out-of-the-box takes a lot of config. But if you're secure out-of-the-box and you offer that sandbox, is that a solid? Strategy,

Ivan Burazin definitely. So especially with the larger companies, so with the like with the smaller younger startups, they don't really care at the small point in time, just like awareness and where the product is and whatnot. But the larger companies definitely, yeah. So it's like. We do like. Pen tests every month we have like. All the. Certifications we have like we can show people and test this. They can. Also there's different deployments where it's like. There's a. Multi cloud, multi tenant cloud, which is our general offering. There's also like a single tenant cloud which gives you kind of completely isolated and there's also completely on Prem. So there's like layers of security, so.

The sandboxes themselves are like, very secure. There's no way to get out. Of that, the whole product itself is. Pen tested, we're completely compliant. The sandbox itself has like. These tools for the. Agent to be secure and then you can also run the whole thing yourself. So there's like multiple layers of security that. We have thought. About that are. Very, very important, as you mentioned, to these large companies, yeah.

Allie Howe Yeah, sounds like a win in terms of just like security and speed statefulness. How does cost for Daytona compared to a Kubernetes setup and a typical hyper scaler?

Ivan Burazin So I think the. Costs are they're. Slightly higher for sure and it also depends on your usage patterns and how much you use and what you. So there's two things we should also mention here is depending on what type of use cases you use Daytona for, your cost will be significantly different. So we have some. Users that have very. Short lived ephemeral workloads and so they will spin up you know of the 1. 8 million like 300,000 a day or 400,000 and it will cost them virtually nothing you know like sub $100 it's like. It's really, really, really. You know, cost effective and then you have others that will spin up, you know, 20,000 a day, but they'll run for hours and hours and so and that will cost them, you know, orders of magnitude more than these ephemeral workloads now comparative to.

What that would have? Cost on a Kubernetes cluster it is more, but it does remove all the headaches. Of your. DevOps team of managing this and all these other things. That being said, comparative to other function as a service. Products it costs less. Than running a Lambda for sure like a single Daytona sandbox cost less than a Lambda. So if you want to equate. To that I would say that it is. Cost for what you get, probably All in all net cheaper than what you would do building it all yourself.

Allie Howe Yeah, I know. Like lambdas have a typical like cold start problem, right? Daytona would not have that.

Ivan Burazin And a statefulness problem, but yes.

Allie Howe Yeah, interesting. One interesting thing I saw, I think maybe you tweeted or said it on a podcast was about like 50% of YC startups today don't have a hyperscaler account. So like an AWS account, for example. Why is that? Like are they mostly just like agent native startups that realize, OK, we need something like Daytona or are they just using other clouds?

Ivan Burazin Or to add a caveat on that like it's not the N isn't the. Entire class there's like. End of, you know, 50 startups. So directionally I feel is correct. And I found that more and more people are aligning with this. And I've heard more like from other podcasts and people researching that in general, this is becoming true. And I don't think it's just there's one thing where, yeah, the younger generation will use, you know, maybe a Vercel instead of, you know, AWS or they'll use render or they'll use a Daytona for the agent part or whatever. Maybe. But also it seems that. The hyperscalers are still.

Growing fairly fast, you know, 10 to 20% year. Over year, great growth. Rough and margins, enormous revenue, but they own a certain segment of the market where you have now these new and we also didn't mention neo clouds, right. So you have like these new neo clouds which are like for inference and whatnot. The CoreWeave, Fireworks of the world. And then you have these new agent clouds, which is like us and other companies. So the cloud. Market is growing faster than the. Legacy clouds are. Growing on themselves and so I don't feel that they are going down in growth. I just feel that the entire.

Market of cloud. If we can call it that is growing faster than they are and they're market ownership will go down. Someone was predicting that AWS will go down to like 7%. Market share of cloud. Which is like extremely small comparative to what they have now. But. They will still be growing in that factor, but just the entire cloud market will be growing so much faster for all these different use cases and they'll be new vendors that will be the winners in these. And so they will take over. And I think that correlates to what we're seeing on a very small scale in these. Like YC type accelerators where there's.

New products and new use cases that need. Specific clouds Which AWS? May have an offering, but it's not a specific offering that will. Give you the best. Or performant or more secure solution to what you need.

Allie Howe Yeah, I think that really plays into something else. I heard you say around like the number of agents in the future or just like, you know, number of humans like to the power of N because every humans just have one agent. We're having more than that. So like we need so much more compute. We need so many, like more new cloud vendors.

Ivan Burazin I keep saying that one because it's the only one that makes sense to me because even today I'm sure I'll let you spun up like 20 agents today or 50 agents today like. Just like every. Chat in ChatGPT is a new agent, let alone all the other. Tools that you use. And if it's trending towards that, every agent will need. A computer to do so. One every agent that you spin up is, you know, running inference, which is, you know, running on a GPU somewhere where that happens. To that website or service or app. Is running somewhere which is like a cloud hosting provider for that and then three, what's more interesting for us like every agent that needs to.

Execute a tasks or a browser code, whatever will be. Spinning up a sandbox and then you multiply. That out as. And looking at the vast majority. Of apps we still. Use are not agents themselves, but once all our apps become agents and once all the agents start doing these things, the number is just like insane, right it? Is literally to the. Power of how many humans there are.

Allie Howe Yeah, amazing. If if we did this podcast again, I'm sure we'll come back and look at 1. 8 million sandboxes a day. Like all time High will be like, you know.

Ivan Burazin Nothing, Nothing. Yeah, Yeah. Nothing, nothing. Yeah. In two months of trajectory where it's going probably like 100 million a day. And so like that's. A scary thought, but. An exciting 1 again.

Allie Howe Yeah, scary. But I'm sure, I'm sure your investors love that number.

Ivan Burazin They love it. Yes, they. Love it. Of course.

Allie Howe What's your advice to like other startup founders that are maybe trying to build agent native companies and are wondering if they're like early or right on time? Because what's interesting to me in the market right now is agents are fundamentally changing so many things about how traditional software works. Like we need these new infra solutions, we need new authentication solutions. So a lot of people are building things right now that maybe they feel like, I don't know, they're early, they're not sure if they're going to be right on time. How do you time the market correctly when the market is evolving so rapidly?

Ivan Burazin It's so hard and we have learned this founding team of Daytona. Also created the very first browser based. IDE in 2009, so like whatever, like almost 20 years ago, not good timing, not good timing like super early, like makes sense today didn't make sense then, right? So timing is important, but what I can? Sort of suggest is. To look at and we're seeing this right now. To look at what people are working. Doing patchwork against. So examples are like when we see agents or agent builders trying to solve their problem and there's no solution on the market. And there's a lot of problems that don't have a solution on the market.

Not all of them weren't creating a solution, right. The thing I can only say is like talk. To or try to build or be part of. Discussion of all these agent building companies and try to find out what they're all consecutively rebuilding and that thing that they're all rebuilding is probably a really good thing to go out and solve. And I think of these things as what I mean specifically is not for humans, but for agents. So what is a new piece of? Software feature tool that. Agents need that the human equivalent is not good enough and you're quite interestingly, I actually put this in some of our talks where we're seeing and I don't think we'll build.

This but what we? Are seeing is that for coding agents which we have a. Lot of they're all. Bypassing and or solving for version control in a different way. So they will use. Git when they're done and they send it back to the human so you can like check it out or. Whatever. But in their inner loop, what they're doing, every single company is doing it differently. Every single one. Like we have one company that basically will dump their entire code base into a JSON file every 5 minutes and that's all it is. Just like to. Dump the whole thing right? Just for speed and efficiency.

And actually today someone sent me a new company trying to solve version. Control for AI agents. And I'm like, this is it. They might have to see, they might, they might not. But the interesting thing is what we're noticing is that this is a problem that's reoccurring and every agent builder is trying to solve that. And that to me seems like a big enough problem to go out and solve. Now, the exact timing is hard, but like if you get the right companies that are building right now and you can offer a solution for them right away, then you become the de facto for that and everyone starts using you as a problem.

So the thing that. I would suggest. And what if I wasn't doing this specifically would be building products for agents instead of humans for sure, just like the market of agent consumption. Well, as far as anything of human consumption. And what are those tools that need to be sort? Of rebuilt or recreated specifically. For agents and they're different enough. That you have to. Find an ICP product.

Allie Howe Yeah, for people that are able to have insight into what those problems are and just figure out like what's a repeatable problem multiple people are experiencing at once, like that's a really great place to be. I want to leave our audience with takeaways or like next steps around, like understanding when they might want to check out Daytona or when they should use a sandbox when they're building their AI agents. Some use cases I think we discussed are maybe like running browser agents, browsing the web when you're doing sort of data analysis, when you're executing code. I wonder also if it's like a good use case for like voice agents for any sort of like latency and lack of speed can sort of create this distrust between a voice agent and the user.

So there's like a giant pause or the voice agent says the wrong thing. Like, that's obviously not a really good example of trustworthy AI. What are some inflection points for those that are building AI agents today for like when they should be checking out Daytona?

Ivan Burazin Depends what they're building and how they're building for sure. But the way I leave people with his first, like most people probably listening are using some personal Claude code on their computer. And I was just like, don't do that. Like from a security standpoint, just don't do that. They turn to give us a bunch of free credit. So like you can probably just like remove that off your machine. I think that would be like the first interaction and when they see. How well Claude code or whatever? They're using but also the ability to spin up like a bunch of them to go out and do the same thing.

I feel that that is. Potentially something that your listeners might be interested in and a sort of a kickoff for that. On the other hand, if they're like a more sophisticated builder and they have a Kubernetes cluster that they're trying to use for these things, please don't do that and take a look at what data does. And so, yeah, I think that would leave them with that.

Allie Howe What's next for Daytona?

Ivan Burazin Oh, so just the composable computer is just missing so many things right now. So right. Now it is a CPU. Linux box and when you think about what we're building if you think about all the types of computers that you can buy as a person and all the. Peripheries that you use. Around your computer and all the software that you have on your computer, a lot of those things are things that we have to build in there, right? So the GPU, different operating systems as we have some of this experimental, but it's not launched yet. So like, there is so much work in giving agents equal rights to computers as the humans have right now, for lack of better word.

Allie Howe Amazing. Well, super excited to keep an eye on Daytona and understand more about how people are using you in the future and come back and maybe revisit where your all time high for number of sandboxes a day is in the future. We'll definitely keep an eye out for more from Ivan and the team. Ivan, where can we will connect with you if they want to join in on the conversation or try Daytona.

Ivan Burazin Yeah, sure, Daytona dot IO is the website, but also like I'm active on like. Twitter or X? LinkedIn and Slack. We have like a pretty. Energetic. Slack, so feel free. To join you can DM me, I'm in there so if anyone has any questions I'm happy to help out.

Allie Howe Amazing. Thank you so much.

Ivan Burazin Thank you, Allie.