Daytona processes 850,000 sandbox runs per day. The company grew 74% month-over-month after pivoting from human developer environments to composable compute for AI agents.
Ivan Burazin, Daytona’s CEO, walked away from $300K ARR to rebuild the company from scratch around a single thesis: AI agents don’t need code execution boxes — they need full computers. The distinction matters. A sandbox runs a script. A “composable computer” gives an agent access to filesystems, networking, package managers, and persistent state.
The numbers from Burazin’s conversation with swyx on Latent Space:
- 60ms cold-start for a single sandbox
- 50,000 concurrent sandboxes spun up in roughly 75 seconds
- Bare metal infrastructure with a custom scheduler — no Kubernetes, no VM overhead
- 74% MoM growth across the platform
- 850K daily runs from agent workloads
Daytona runs on bare metal because the latency math doesn’t work otherwise. When an agent needs to install a dependency, compile code, or test a network call mid-reasoning, a 2-second container spin-up breaks the loop. Sixty milliseconds doesn’t. Burazin’s framing — “composable computers for agents” — positions Daytona alongside the small cluster of AI infra companies (Modal, E2B, Fly.io) building the runtime layer beneath agent frameworks.
The business model shift is worth noting separately. Daytona moved from per-seat SaaS pricing to consumption-based billing — a pattern that tracks with how agent workloads actually scale. One customer running 10,000 agent tasks a day generates more value than 50 developers with monthly seats.
NOTE
Burazin also discussed using reinforcement learning for evals — treating sandbox outcomes as reward signals for agent training loops. This is infrastructure-level work, not application-layer prompt tuning.
Daytona sits in the “agent cloud” tier that barely existed 18 months ago. The 74% MoM growth rate suggests agent compute demand is compounding faster than most teams’ infrastructure can keep up.