Meet OSGym: A New OS Infrastructure Framework That Manages 1,000+ Replicas at $0.23/Day for Computer Use Agent Research

**OSGym** is a new open-source infrastructure framework from MIT, UIUC, CMU, and UC Berkeley that dramatically reduces the cost of training computer use agents by up to **90%**, managing over **1,000 virtual desktop environments** simultaneously for as little as **$0.23 per replica per day**.

The framework addresses the major infrastructure challenge of training AI agents that can operate computers through actual GUI interactions by implementing four key optimizations:

• **Decentralized state management** - Each OS replica has its own dedicated manager to prevent single points of failure • **Hardware-aware orchestration** - Shifts bottleneck from expensive CPU to cheaper RAM by packing more replicas per server • **Copy-on-write disk management** - Reduces storage consumption by **88%** and speeds VM provisioning by **37x** using XFS reflink technology • **Robust container pooling** - Maintains pre-warmed VM pools with multi-layer fault recovery

In practice, OSGym collected **1,420 trajectories per minute** using 1,024 parallel replicas, with the entire dataset costing just **$43 in cloud compute**, making computer use agent research financially viable for academic budgets.