AI Agents Need Computers: 74% MoM Growth, 850K/Day Runs, & New Agent Cloud — Ivan Burazin, Daytona
Daytona CEO Ivan Burazin discusses the company's pivot from developer environments to composable computers for AI agents. He explains the unique infrastructure challenges posed by spiky RL and eval workloads, Daytona's bare-metal architecture with a custom scheduler for high performance, and the future need for stateful Windows and macOS sandboxes to automate knowledge work.
The journey to building infrastructure for AI agents began with the thesis that “localhost should die.” This idea initially led to CodeAnywhere, one of the first browser-based IDEs, and later to Daytona, which first focused on automating development environments for human engineers. However, the team quickly realized that the emerging market of AI agents had fundamentally different needs than human developers.
The Pivot to AI Sandboxes
At the end of 2023, while experimenting with open-source agent frameworks, the Daytona team was approached by numerous agent developers. They needed a sandboxed compute environment for their agents, but the existing Daytona product, built for humans, was not suitable. After receiving consistent feedback that their current infrastructure “breaks” for agentic workloads, the team went back to the drawing board.
Over New Year’s Eve, a new, minimalist MVP was created, designed from first principles for agents. When demoed to the same developers who had rejected the previous version, every call went long, and every developer demanded immediate access. This intense, unsolicited pull from the market signaled a clear product-market fit and a massive, untapped opportunity: providing the fundamental compute layer for every AI agent. This led to a hard pivot, focusing the entire company on building “composable computers for AI agents.”
Architecture: Bare Metal, Statefulness, and Speed
The core insight was that agents require a combination of the speed of serverless functions and the statefulness of a persistent VM. They need environments that can be spun up instantly but also paused and resumed, just like a human closes and opens a laptop lid. To achieve this, Daytona built its infrastructure on key principles that differ from conventional cloud offerings:
- Bare Metal and a Custom Scheduler: Instead of running on top of VMs and Kubernetes, Daytona runs on bare metal with its own custom scheduler, a spiritual successor to the one built for CodeAnywhere. This approach eliminates layers of abstraction, allowing for direct access to the underlying machine’s CPU, RAM, and NVMe drives.
- Local Snapshots for Instant Startup: Templates and snapshots are preloaded onto the bare-metal machines. When a new sandbox is requested, the scheduler directs the request to a machine where the required snapshot already exists locally. This eliminates network latency associated with fetching images or attaching network-based storage (like EBS), enabling a startup time of approximately 60 milliseconds.
- High Concurrency and Scale: This architecture allows for massive concurrency. Daytona can spin up 50,000 sandboxes in about 75 seconds, a task that could take 30 minutes on other platforms. Their largest customer runs nearly 850,000 sandboxes daily, with requests for up to 500,000 concurrent CPUs.
- Dynamic Resizing: A key feature is the ability to dynamically resize a sandbox’s resources on the fly, making it extremely difficult to cause an “out of memory” (OOM) error. This is a significant advantage over platforms where resources are fixed upon creation.
The New Infrastructure Problem: Spiky Workloads
The usage patterns of AI agents are drastically different from human users. Daytona identifies two primary workload types:
- Long-Running Background Agents: These agents (e.g., coding assistants) exhibit usage patterns similar to humans, often described as “follow-the-sun,” with peaks during the workday and lulls at night and on weekends.
- Reinforcement Learning (RL) and Evals: This workload, which grew from 0% to nearly 50% of Daytona’s usage in a few months, is extremely spiky. A user will go from zero usage to consuming 100,000 CPUs for a run, hold it at 100% utilization, and then drop back to zero.
This spikiness creates a new infrastructure challenge. To serve these bursts, providers must have capacity available for peak demand, leading to very low mean utilization—Daytona’s is around 15%, despite hitting 90% during peaks. This is a shared problem across the AI infrastructure space, from databases to compute providers. Customers running these workloads often migrate from managed Kubernetes (EKS, GKS), citing Daytona’s superior ergonomics, speed, and scale as reasons they are “never going back.”
The Future: Composable Computers and Legacy Systems
The vision extends beyond simple Linux environments. The belief is that every agent will eventually need a “computer,” a fully-featured environment tailored to its task. This is crucial for unlocking the massive market of automating knowledge work, much of which is trapped in legacy applications.
- Windows and macOS Sandboxes: To address this, Daytona is developing full-fledged Windows and macOS sandboxes. This allows agents to operate in the same environments as human knowledge workers, interacting with applications that lack modern APIs. An agent can be instructed to log into a website, export a CSV, and perform analysis—tasks that are impossible in a headless Linux environment.
- Licensing Challenges: Providing macOS at scale is particularly difficult due to Apple’s restrictive licensing. Current rules only allow two VMs per physical machine, and a license can only be reassigned to a new user every 24 hours. Furthermore, memory snapshots can only be resumed on the same physical machine, preventing the load balancing and workload migration that is standard for Linux and Windows.
Ultimately, the future AI cloud may look less like AWS and more like Stripe—a set of powerful, API-first primitives built specifically for agentic consumption. This includes not just compute sandboxes but also agent-native databases, web search tools, and other primitives that are yet to be invented.