Ivan Burazin

Ivan Burazin is a co-founder of Daytona and led its January 2025 pivot from development environments for human engineers to compute sandboxes for AI agents. In his May 2026 Latent Space interview, he described the goal as “composable computers for AI agents”: computers accessible through an API, with persistent state and configurations suited to different tasks, rather than merely boxes that execute code and return an output.[1][3]

The change began with a late-2024 experiment combining Daytona’s development-environment infrastructure with OpenDevin, later called OpenHands, as a hosted service. Burazin recalled that few people used it, but agent builders approached the company seeking compute runtimes. Giving them the existing infrastructure exposed a mismatch: roughly 20–30 prospective users rejected it. The founders concluded that agents needed a different compute primitive, although they had not yet worked out its form.[1]

Burazin built the first prototype on New Year’s Eve 2024, describing it in his interview published in May 2026 as partly “vibe coded.” His CTO rejected the implementation but endorsed the idea and rebuilt it over two weeks. When they demonstrated the replacement to people who had rejected the earlier product, Burazin recalled that calls ran long and users followed up asking, “Where is my API key?” That response convinced the founders to commit to the sandbox product. The first version was a Linux computer with configurable CPU, disk and RAM; the broader vision of multiple operating systems, GPUs and live resizing came later.[1]

Burazin’s design premise was that an agent should be able to pause and return to the same computer state, much as a person closes and reopens a laptop. He said Daytona combined that persistence with fast startup by running on bare metal with its own scheduler. After evaluating Kubernetes and Nomad against their requirements, his CTO drew on scheduler experience from Codeanywhere. Templates and snapshots were preloaded on local NVMe drives, allowing requests to be routed to a machine already holding the required snapshot.[1]

In the interview, Burazin reported 60 milliseconds for one sandbox’s request-to-reply startup, including network latency, and about 75 seconds to start 50,000 sandboxes concurrently. Separately, he said Daytona’s largest customer ran about 850,000 sandboxes per day. These were company-reported figures for distinct measures: individual startup latency, concurrent startup time and one customer’s daily volume.[1]

He distinguished background agents, whose demand followed human working hours, from reinforcement-learning and evaluation workloads that could jump abruptly from idle to very large runs. Daytona sought customer commitments to reserve capacity for those peaks. Burazin explained the trade-off through GPU utilization: provisioning CPU environments on demand could leave more expensive GPUs waiting between training cycles. For repeated runs, he described building and propagating dependency snapshots; he also described live resizing and Docker hardened with Sysbox, supporting nested Docker and K3S workloads.[1]

Burazin also drew a boundary around infrastructure ownership. Daytona ran bare-metal machines through colocation providers, but he said owning data centers did not then justify the capital and risk for an expected margin improvement of only single-digit percentage points. The architecture allowed that option without making it an immediate business priority.[1]

Although Burazin preferred headless or terminal operation where possible, he backed substantial investment in computer use. His own board-report workflow illustrated why: integrations returned polished reports with missing data, so he gave an OpenClaw agent in a Mac Mini virtual sandbox its own company identity and read-only service accounts. He recalled directing it to log into websites and export information unavailable through the integrations. Daytona’s computer-use offering was not generally available in that account, and he described macOS sandboxes as forthcoming.[1]

Burazin said he personally joined many sales calls and customer Slack channels, using repeated requests to guide the roadmap. He regarded team responsiveness as a reason customers returned, alongside speed, capacity and features. He connected that working culture to longstanding collaborators: of Daytona’s 25 people at the time, he estimated that about 13 had worked with the founders for more than seven years.[1]

Burazin described adopting AGPL 3 for the sandbox product to retain Daytona’s open-source community while adding restrictions, replacing the earlier arrangement of a permissively licensed core and proprietary enterprise packaging. He said repository access gave agents more context when integrating Daytona, but did not significantly drive growth at that time. His account excluded feature-flagged work such as Windows and GPU capabilities from the open-source offering.[1]

He saw room for agent-oriented versioning before code reaches the usual GitHub collaboration workflow. One customer, he recalled, repeatedly serialized an entire sandbox codebase into JSON and stored it in S3 because that was fast enough for its needs. Burazin expected Git, and possibly GitHub, to persist while new tools served agents’ inner development loop. He also discussed sandboxes as possible CI runner infrastructure, while explicitly saying Daytona was not a CI provider.[1]

Burazin argued that SaaS vendors should expose their data to customers’ agents and charge for API consumption: “Just expose everything and charge me for that.” He questioned treating revenue from resold model tokens as equivalent to traditional SaaS revenue, citing different margins and the persistence of data silos. His broader forecast was an agent-specific cloud combining sandboxes, search, databases and additional primitives, with consumption resembling Stripe more than AWS.[1]

Before the agent pivot, Burazin had worked on cloud development tools and developer communities. He co-founded Codeanywhere with Vedran Jukić, founded the Shift developer conference and subsequently led developer experience at Infobip after it acquired Shift. Looking back in 2026, Burazin credited his co-founder with the browser-based IDE’s invention and said building its underlying stack supplied technical experience later used at Daytona.[2][1]

On September 1, 2023, Burazin left Infobip’s developer-experience leadership to focus on Daytona with the original Codeanywhere team. The co-founders’ roles were CEO for Burazin, CTO for Vedran Jukić and chief architect for Goran Draganić. At that transition, he was to remain an adviser to Shift’s organization, while Nikola Pavešić took over developer experience.[3]

In a September 2023 interview, Burazin said another developer-tool company had not initially been his plan; interest from analysts, investors and customers persuaded him the timing was right. He distinguished Codeanywhere’s browser IDE interface from Daytona’s infrastructure orchestration, which divided company infrastructure into developer workspaces and managed environments within existing workflows. His stated preference was to improve tools developers already loved rather than require new ones. That was the human-development business from which the later agent-sandbox pivot departed.[2][1]