Training computer-use agents (CUAs) at scale is bottlenecked by the scarcity of diverse, reliable software environments. Native application environments demand heavy construction and integration work, which restricts coverage and risks overfitting to specific applications, while browser-hosted environments offer limited interaction depth and workflow complexity.
To address this, researchers have proposed SoftwareGym, an automated pipeline that synthesizes GUI software environments with diverse interfaces, executable functionality, and runtime-verified behavior. The system generates structurally varied designs, realizes them as runnable applications, and applies GUI-based evaluation with iterative repair to ensure functional reliability.
Using this pipeline, the team synthesized 182 applications across twelve categories, achieving a mean GUI task pass rate of 91.5%. They also built SoftwareGym-Bench, a 120-task evaluation suite spanning the same twelve categories, and collected trajectories from these environments to train a CUA model.
The resulting model improved Pass@3 from 41.39% to 56.97% on the strictly out-of-domain OSWorld benchmark and reached 40.0% on SoftwareGym-Bench. According to the authors, these results demonstrate the potential of synthesized software environments as scalable and transferable supervision for CUA training.