OSWorld on Daytona Windows Sandboxes

OSWorld serves as a prominent benchmark for evaluating computer-use agents by giving them natural language instructions to perform tasks on live desktops using screenshots, a mouse, and a keyboard, with a script inspecting and scoring their attempts. While the benchmark includes tasks for both Ubuntu and Windows operating systems, the latter requires setting up a virtual environment manually, as there is no pre-packaged image available. The piece details an experiment where Windows tasks were executed using Daytona Windows sandboxes, focusing on the performance of Claude Opus 4.6 and GPT-5.4 agents. These agents were tasked with activities ranging from single-app tasks in Microsoft Office to complex multi-app workflows, with results showing varied success rates across different categories. The document highlights the process of setting up the environment, running tasks, and the modifications made to evaluators to address scoring issues, further providing a comprehensive guide for replicating the experiment.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.