Giving Agents Computers — Ivan Burazin, Daytona

Episode Summary

Executive Summary: Ivan Burazin traces Daytona’s evolution from Code Anywhere and Shift to a bare-metal, open-source compute platform for AI agents. The conversation centers on why agents need fast, stateful, composable computers; how spiky AI workloads reshape infrastructure economics; and why computer use, Windows, and eventually GPU sandboxes could become major agent primitives.

Main Topics: Origin story: Code Anywhere to Daytona (Priority: 5/5): Burazin explains the long arc from early browser-based IDE work to Shift conferences and back into infra via Daytona, emphasizing repeated pivots with his co-founder and lessons from building the first browser IDE. Why Daytona exists: computers for agents (Priority: 5/5): Daytona is framed not as a generic sandbox but as composable computers for AI agents—fast, stateful, and adaptable to different workloads like terminals, Windows desktops, and RL runs. Bare metal architecture and performance (Priority: 5/5): The team chose bare metal, custom scheduling, local snapshots, and multiple isolation layers to deliver millisecond startup times, high concurrency, and the ability to resize and resume workloads dynamically. Workload shift: background agents and RL/evals (Priority: 4/5): Burazin distinguishes follow-the-sun background agents from spiky research/RL workloads, noting that new usage patterns force a rethink of provisioning, commitments, and capacity management. Computer use and legacy software (Priority: 5/5): A major bet is computer-use sandboxes for Windows, Mac, and Linux, aimed at automating legacy apps that lack complete APIs and unlocking broader enterprise and knowledge-work automation. Open source, GTM, and enterprise responsiveness (Priority: 4/5): Daytona uses open source as part of product adoption and context-sharing, but Burazin says the real sales differentiator is fast, high-trust support and responsiveness across Slack and live calls. Market outlook and infrastructure thesis (Priority: 5/5): Burazin argues that agent infrastructure is becoming its own cloud stack, with sandboxes, search, databases, and more primitives still to be built; he also warns that token-reselling SaaS may be overvalued.

Key Arguments: Agents need stateful, resumable computers, not just ephemeral containers, because their work resembles human laptop usage more than batch jobs. Bare metal plus a custom scheduler enables Daytona’s speed and local snapshotting, avoiding the latency and constraints of VM-based stacks. The market pivot happened when customers building agents repeatedly asked for the same primitive Daytona had begun to build, validating the sandbox thesis. There are materially different workload shapes: background agents behave like human usage, while RL/evals are extremely spiky and require different capacity planning. Computer use is necessary even when APIs exist, because many enterprise workflows still require logging into legacy UIs and exporting or manipulating data manually. Open source helps adoption and customer context, but enterprise conversion is driven more by support quality, trust, and product performance than by GitHub stars. The agent economy will likely demand a broader infrastructure cloud, with sandboxes as one primitive among many rather than the final form of the stack. Current SaaS market narratives overstate token-based revenue re-acceleration; real value comes from exposing APIs and charging for true consumption.

Data Points: Code Anywhere users: about 3 million - Burazin says the original browser-based IDE reached roughly 3 million users. Daytona team size: 25 - He says Daytona has 25 people today. Long-tenured teammates: about 13 of 25 - Roughly half the team worked with Burazin for seven-plus years. Reported growth: 74% month-on-month - The podcast references Daytona’s rapid growth during the market expansion. Spin-up time for one sandbox: 60 milliseconds - Daytona claims request-to-run-response latency including network latency. Concurrent spin-up of 50,000 sandboxes: about 75 seconds - Burazin contrasts Daytona’s scale with slower competitors taking much longer. Slow competitor benchmark: about 2,000 seconds - He cites public data for some providers taking around 30 minutes. Biggest customer daily volume: about 850,000 to just under 1,000,000 per day - He describes one customer running sandboxes at very high daily frequency. Requested concurrency: 500,000 concurrent - A customer requested half a million concurrent CPUs/sandboxes. Mean utilization: 1.5% - Daytona’s utilization is low because workloads are highly spiky. Peak utilization: up to 90% - Specific workloads can briefly drive utilization near full capacity. Number of users/skills by region: Singapore is number one city by users - Burazin notes surprising geographic adoption patterns. Knowledge workers in US: about 100 million - Used to size the computer-use and automation market. Knowledge workers worldwide: about 1 billion - Used in the market-sizing argument for agentic automation. Aggregate wages in US knowledge work: $10 trillion - Part of the argument that computer-use automation could address huge economic value. Aggregate wages worldwide: $50 trillion - Supports the scale thesis for agent-driven work automation. Open source license: AGPLv3 - Daytona’s sandbox product is open source under AGPLv3 for reciprocity. Mac OS parallel VM limit: 2 per machine - Burazin explains a licensing constraint that complicates Mac sandboxes.

Pivotal Quotes: "What Daytona does today is essentially composable computers for AI agents." — Ivan Burazin: Defines the company’s core product beyond the “run AI code” shirt slogan. "I've never felt so... people literally call you if you do not give them access. Like they want access right now." — Ivan Burazin: Describing the market pull after showing the improved sandbox prototype to customers. "I don't want your agent essentially because... just expose everything and charge me for that." — Ivan Burazin: His critique of token-wrapped SaaS and preference for direct API consumption.

Implications: AI infra is shifting from generic cloud primitives to agent-specific computers, with speed, state, and computer-use support becoming strategic moats. Enterprises will likely adopt open, responsive infrastructure vendors that can handle spiky workloads and legacy UIs.