Authors: Elena Vrabie and Adriana Spulber

"Can I spin up 500 environments in parallel?" When Ivan Burazin, co-founder and CEO of Daytona, started hearing questions like this from customers, he knew something was shifting. The Croatian startup had built cloud development environments for human developers, but their users weren't human anymore. They were AI agents. 

Ivan Burazin is a serial entrepreneur with a background in engineering who previously co-founded Codeanywhere and the Shift Conference, which got acquired by Infobip (Croatia’s first unicorn), and where he later became Chief Developer Experience Officer. Known for his risk-taking approach, he's now leading Daytona through a hypergrowth AI wave.

Daytona provides secure, elastic infrastructure for AI agents to execute code autonomously. Originally launched as an open-source tool to standardize development environments, the company pivoted in 2024 to focus entirely on agent runtimes. Now having over €31M in total venture capital funding, the company is building what Ivan called in previous talks "a composable computer for an AI agent" - instant, isolated environments that let agents work without human supervision.

In this interview, Ivan discusses Daytona's pivot after customers started requesting features no human would need, and what "agentic AI infrastructure" actually means in practice. He covers sub-60ms sandbox creation, deployment options, production-scale usage reaching millions of daily sandboxes, and how the team balances open-source community building with commercial monetization.

UV: You hit $1M in ARR in 2 months, and social media is on fire with customer case studies. Let's rewind: What made you realize Daytona needed to pivot a little over a year ago from dev environments to agent infrastructure?

Ivan Burazin: The pivot didn’t start with a thesis; it started with customers doing weird things. We built Daytona originally for humans: cloud dev environments, secure sandboxes, fast startup, all the usual DX (digital experience) problems. 

Then customers started asking questions that didn’t make sense in a human world.

• “Can I spin up 500 environments in parallel?”

• “Can I snapshot mid execution and branch?”

• “Can this thing keep running for 24 hours without anyone attached?”

At first, we assumed it was just power users. Then we realized there was no human at all.

UV: The term "agentic AI infrastructure" means different things to people. How do you define it? Where does Daytona fit in the stack?

IB: Agentic AI infrastructure is everything required to let an agent do real work without a human babysitting it.

Daytona sits at the runtime layer. We give every agent a real, isolated, stateful computer that can be created in milliseconds and run for minutes, hours, or days. Above us are agents and frameworks. Below us is bare metal.

UV: Many enterprise conversations center around BYOC (Bring Your Own Cloud). What options does Daytona offer teams that want control over where their agent infrastructure runs?

IB: This is true, but Daytona has multiple deployment options:

1. Our own multi-tenant cloud;

2. A single-tenant cloud;

3. Customer-managed compute, where the control plane is with Daytona, but the runtimes are in the customer;

4. Fully on-prem, where the customer can run it on their own servers;

5. Can run in things like EKS (​​Elastic Kubernetes Service) inside their own cloud.

UV: When you scaled to $1M ARR in 60 days, where did the cracks show up - core product or the edges?

IB: We had many growing pains as Daytona grew over the months. Every time, it was a different issue, and the things that "break" are usually not the things that you expect. They are not the core elements of the product, but supporting services that, unfortunately, did not scale well with the usage

UV: The RPA era taught us that scaling bots is hard. What's Daytona's role when enterprises start running hundreds of agents in production?

IB: We don't necessarily solve for the agent's problem, but we enable multiple agents to solve their problem. That being said, as agents become more sophisticated, we see the rise of Claude Code and cloud bots doing things much, much better. Each of those needs its own runtime, and Daytona is the one that powers that.

UV: Daytona's sub-60ms sandbox creation is what everybody's talking about on the forums. What makes that possible?

IB: The reason we can do that is that we created a stack all the way down to the bare metal and optimized everything specifically for agents' runtimes, and because of that, we were able to do very specific things, such as being very fast.

UV: Agent use cases are evolving fast. How do you decide what to build when customer needs are changing daily?

IB: We're actually surprised every day, mostly because the agent's needs have not been defined and have changed every single day.

What we do at Daytona is we constantly listen to and track questions and feedback from our customers. As soon as we have at least three asking for a similar thing, we then try to verify that internally. 

If it makes sense, we can create that feature quite fast, so that if other companies need that, we are the ones who can provide that specifically. Also, we have been wrong, and we created features that only one or two companies used that we're now going to deprecate. So it is very much about adapting very fast to this new market and making sure you have the features that are actually needed for agents.

UV: Most companies are still in pilot mode. What does 'production scale' actually look like for your customers?

IB: We have companies from YC to Fortune 100. Obviously, some companies are still playing around, but we have companies at massive scales spinning up millions of sandboxes every day, and you're constantly surprised at which ones those are.

UV: Are regulated industries demanding self-hosted Daytona, and does that change your unit economics? Also, your GitHub launch drove community traction, too. 

IB: Not just regulated industries, but any public companies or high-security companies are actually demanding a self-hosted version. It doesn't change our unit economics much; it actually works in our favor, but it does have the overhead of support and upgrading.

I can say that almost every company that does this at scale comes to us for a commercial license or the cloud because running this at scale is not a trivial task.