microsoft/WindowsAgentArena ? reverse-engineered prompt

Build me a Windows AI agent testing setup that can run desktop tasks inside a Windows 11 virtual machine and let me benchmark how well an AI can use the computer.

I want it to work locally with Docker and Python, and I should be able to give it an OpenAI or Azure OpenAI key, prepare the Windows image once, then run lots of tasks and see results quickly. Please include the pieces needed to start the environment, launch the VM, connect an agent to it, and run benchmark tasks at scale. A simple way to view progress and logs would be great too.

Make the default experience easy to follow from the README, and if anything about the latest setup or API usage is unclear, look up the current docs online and fill in the gaps. The goal is a reproducible Windows agent arena I can use to test desktop automation and compare different agent approaches.