New Post: New Investment: Daytona, Building the Computer for AI Agents
Every human worker needs a computer. In the next decade, every AI agent will too — except agents won’t need one computer. They’ll need millions: instantly provisioned, securely isolated, and spun up in parallel as they work.
I’m excited to share that FirstMark led Daytona’s $24M Series A, and I’m joining the board. The round includes Pace Capital, existing investors Upfront Ventures and E2VC, and strategic investments from Datadog and Figma Ventures.
The cloud we have wasn’t built for agents
The cloud we have today was built to ship software: stateless services, immutable deployments, reliable APIs. That model is great for the “outer loop” of production code. It’s not how work actually happens. Agents, like humans, iterate. They install dependencies. They write files. They run tests. They debug. They make mistakes, branch, recover, and try again — at machine speed, and across many parallel threads.
When you force that into today’s primitives, you hit a wall. VMs are heavy and expensive to keep warm. Serverless is elastic but fundamentally stateless, pushing teams into constant state hydration and orchestration complexity. Agents need a third paradigm: instant, stateful, ephemeral compute. Something that feels less like a disposable sandbox and more like a programmable computer.
Daytona’s core insight: the agent “computer” is a new primitive
That’s Daytona’s core idea. The word “sandbox” undersells what they’re building. A computer is a composition: CPU, memory, disk, networking, isolation boundaries, an operating system, and the tooling that makes real work possible. Daytona makes that composition available programmatically. An agent can spin up an environment in milliseconds, keep state across steps, snapshot and fork execution paths, and run reliably for minutes, hours, or days. Done right, this becomes a new infrastructure primitive for the agentic economy.
We led because we believe the category is real and enormous. If agents become a mainstream labor force, “agent compute” becomes one of the largest new markets in infrastructure, with a scale that ultimately dwarfs human-centric computing. And we think Daytona is building the right abstraction early, with a team that has unusually strong founder-market fit. Ivan Burazin and Vedran Jukić have spent years building developer tools and cloud environments, and they’ve made the hard calls required to focus Daytona on agents rather than the more comfortable path of building yet another dev environment product.
What comes next
Building “computers for agents” is deceptively hard. Plenty of large software platforms will ship “a sandbox,” and plenty of teams will try to build their own internally. But once you need instant startup, real statefulness, strong isolation, and production-grade reliability at scale, it becomes a full-stack problem, where every component needs to be designed natively — and single-minded focus matters.
The surface area is enormous: hard security boundaries,...