Daytona Introduces Composable AI Sandboxes, Bare‑Metal Compute

Daytona offers composable, stateful sandboxes that can be accessed through an API, replacing traditional local development environments for AI agents. The platform achieves a startup time of roughly 60 ms and can spin up 50 000 sandboxes in about 75 seconds, with its largest customer running approximately 850 000 sandboxes per day. Daytona’s infrastructure runs on bare‑metal servers with its own scheduler, supporting Windows and macOS environments to accommodate diverse agent workloads. Reinforcement‑learning and evaluation workloads now account for roughly 50 % of Daytona’s usage, creating significant CPU spikes that the platform handles with dynamic resource allocation. Daytona positions itself as an AI cloud that resembles Stripe more than AWS, focusing on API‑first business models and token resale considerations. The company emphasizes the importance of CLI over managed control plane (MCP) for agent power and integration. Daytona’s approach addresses the need for instant startup, dynamic scaling, and composability in AI agent workflows. The platform’s design also highlights the challenges of licensing macOS sandboxes and the benefits of open‑source integration.