The Infrastructure of Agency: Why AI Needs Composable Computers

The core idea here is that AI agents are not just software processes. They are autonomous knowledge workers that require a new approach to compute. Ivan Burazin, CEO of Daytona, argues that the localhost model is outdated because agents need stateful, instant-start, and composable environments that reflect the way humans use computers. The result of this shift is that the AI cloud will not look like the centralized, monolithic AWS of the past. Instead, it will be a fragmented, API-first ecosystem where agents consume infrastructure primitives programmatically. Those who understand this transition gain an advantage: they stop building for simple code execution and start building for agentic workflows, avoiding the bottlenecks that will hinder teams relying on standard, static Kubernetes clusters.

The Hidden Cost of Fast Solutions

Most teams try to build agent infrastructure using standard Kubernetes or ephemeral containers. Burazin argues this is a mistake that creates unnecessary complexity. While these tools handle isolation, they do not provide the statefulness required for long-running agents.

The insight that made us do this pivot is... people did not like what we had built. It did not work... We saw that everyone we gave it to, it was like 20, 30 people, they all said, 'No.' Like, 'This is not what we need. This sort of breaks.'

-- Ivan Burazin

The system dynamics are clear: developers focus on the pain of initial setup and ignore the operational maintenance that grows over time. Daytona’s move to bare metal with a custom scheduler solves this by prioritizing IOPS and stateful snapshots, allowing agents to pause and resume work just as a human would with a laptop.

The 18-Month Payoff: Why Agents Need Computers

Conventional wisdom suggests agents should be headless and interact only via APIs. Burazin points out a non-obvious dynamic: the knowledge work market is stuck in legacy Windows and macOS applications that lack sufficient API coverage.

• Immediate Benefit: Headless agents handle simple data extraction.
• Hidden Cost: When the agent hits a siloed legacy app, it fails or provides partial data, which forces manual human intervention.
• Lasting Advantage: Giving agents a composable computer, such as a virtualized Windows or macOS environment, allows them to navigate legacy UIs, export data, and finish tasks end-to-end.

This approach requires significant upfront effort to manage licensing and virtualized OS overhead, which most teams avoid. That difficulty is exactly what creates a competitive moat.

How the System Responds to Spiky Workloads

The most important systems-thinking insight from the conversation is the shift in usage patterns. Traditional human development environments follow a predictable pattern. Agentic reinforcement learning and evaluation workloads, however, are violently spiky.

When you have companies doing sort of like evals and RL, they are super spiky. So they are gonna come in, it is like, 'We are gonna use nothing, then can we have 100,000?' Right? And then go back down.

-- Ivan Burazin

This creates a feedback loop where traditional over-provisioning becomes too expensive. The system forces infrastructure providers to either build just-in-time compute, which introduces latency that hurts GPU utilization, or to own the bare metal to ensure the instant-on capability required to keep expensive GPUs fed.

Key Action Items

• Audit your agent infrastructure for statefulness: If your agents require a re-run from scratch because they lose state, you are paying a reboot tax that will grow as your agent complexity increases. (Immediate)
• Shift from code execution to computer provisioning: Stop viewing sandboxes as disposable containers. Over the next quarter, evaluate whether your agents need OS-level access to navigate legacy UIs that lack APIs. (Over the next quarter)
• Prepare for spiky economics: If you are building agentic workflows, move away from auto-scaling groups that take minutes to start. You need infrastructure that can handle 100,000 CPU spikes in seconds to maintain GPU efficiency. (12-18 months)
• Expose data via APIs: Stop reselling tokens as a SaaS feature. The lasting advantage lies in exposing your product's internal data via APIs so agents can consume it programmatically, rather than forcing them to scrape your UI. (6-12 months)
• Prioritize responsiveness in vendor selection: As the agentic market matures, the technical difference between providers will narrow. The durable advantage is working with vendors who provide direct, Huddle-based support for operational issues. (Immediate)