Daytona’s CEO, Ivan Burazin, recently elaborated on the company’s significant shift from traditional human development environments to specialized agent sandboxes designed for AI. This pivot responds to the growing demand for dynamic, stateful infrastructure capable of supporting unpredictable workloads, particularly in reinforcement learning and evaluation tasks, which have rapidly increased to account for nearly 50% of Daytona’s usage. As agent compute providers enhance their offerings with composable computer environments—prioritizing API-first accessibility—they cement their position in the evolving market, where seamless scalability and real-time responsiveness have become essential for effectively leveraging AI agents.

Manus: Manus is an AI agent platform that equips its agents with computer access for complex tasks. It exemplifies the growing demand for agent-native compute infrastructure that supports full, stateful environments beyond basic code execution.
Cursor: Cursor is an AI-powered code editor that incorporates computer-like capabilities for agents. It illustrates the consolidation of developer tools around agent-first computing environments in the current LLM OS ecosystem.
GDPVal: GDPVal is an agentic evaluation benchmark that relies on computer environments for agent assessments. It reflects the broader shift in research tools toward assuming composable, stateful compute for realistic agent benchmarking.
Harbor: Harbor is a tool or framework in the agentic evaluation space that assumes computer access for agents. It plays a role in standardizing computer use across research benchmarks and supports the emerging needs of agent infrastructure.
Daytona: Daytona provides composable computers and stateful sandboxes for AI agents via API access, supporting dynamic resources, multiple operating systems, and long-running workloads. The company recently pivoted from human developer environments to agent-native infrastructure built on bare metal with custom scheduling for instant startup and high concurrency. This shift addresses the needs of agents that require production-grade, stateful environments instead of disposable code execution boxes.
Perplexity: Perplexity is an AI search platform that has adopted computer use capabilities to enhance its agent operations. It represents the product-side trend of AI companies integrating composable computing environments as part of the evolving LLM OS stack for more capable agent workflows.
Ivan Burazin: Ivan Burazin is the CEO and co-founder of Daytona with prior experience building early cloud development tools including CodeAnywhere. He has long advocated for moving development away from fragile local machines and now applies that vision to AI agents needing reliable, API-accessible computers. In recent discussions, Burazin details the market pivot toward agent-first sandboxes and the infrastructure requirements for scaling reinforcement learning and evaluation workloads.
TerminalBench: TerminalBench is an agentic evaluation benchmark that assumes access to computer environments for testing agent performance. It contributes to the research-side standardization of computer use in agent evaluations and workflows.

`json
{
“Market Positioning”: “Agent compute providers are positioning themselves as specialized infrastructure for AI agents, emphasizing API-first models and dynamic scalability over general-purpose cloud offerings.”,
“Workload Evolution”: “Reinforcement learning and evaluation workloads now drive significant usage of agent compute platforms due to their unpredictable, high-concurrency patterns that differ from traditional developer workloads.”,
“Agent Infrastructure Trends”: “Product and research efforts in AI are converging on composable computer environments as a core toolkit for agents, with platforms adopting full computer access rather than limited execution sandboxes.”
}
`