Francesco Bonacci is a founder of Cua, which provides desktop environments and tools for computer-use agents. He is a co-author of Windows Agent Arena, a benchmark that evaluates agents inside real Windows applications rather than limiting them to browser or text-only tasks.[1][3][4]
Start with Windows Agent Arena →
Links
Founder profile
GitHub 
X Founder · Cua
Career
Bonacci’s founder profile records previous work at Xbox and Microsoft AI and identifies him with Cua, founded in 2025. Cua packages computers as environments that agents can control through screenshots and input tools. The company’s launch description connects virtual machines with an agent-friendly interface so developers can test desktop automation without giving a model unrestricted control of their everyday machine.[1][4]
Contributions
Bonacci co-authored the 2024 Windows Agent Arena paper. The team adapted OSWorld to more than 150 Windows tasks requiring planning, screen understanding, and tool use. Parallel evaluation on Azure reduced the time needed to run the suite. Its Navi agent completed 19.5% of the Windows tasks, compared with 74.5% for unassisted humans in that study. Those figures describe a particular benchmark and agent, while the lasting contribution is a reproducible environment for studying where desktop agents fail.[3]
Ideas
Cua’s open-source repository covers drivers, isolated operating-system environments, and benchmarks. These pieces address different stages of computer use: getting observations and actions to an application, supplying a machine on which to run it, and checking whether the task succeeded. Bonacci’s connection to the project is as a founder; the broader repository is team work rather than a claim that he individually authored every component.[4][5]
Tools, writing & talks
Writing, research & talks
Start here
Windows Agent Arenaarxiv.orgA co-authored benchmark and agent study using real Windows applications and parallel evaluations.
Begin with the abstract and task examples to see what desktop agents are actually asked to do. The reported gap to humans gives a concrete reason to care about infrastructure and evaluation before exploring the code.
Co-author[3]Cua explains how jev-use turns screen state into bounded action candidates that a fast decision model (Jev) scores instead of calling a general LLM, how OmniParser fills the perception gap for screenshot-only UIs, and where the 'System 1/2' framing for small models breaks down.
Catalog workOriginal sourceCua demoed a technical preview of background computer use on Wayland. It tested two Hyprland-oriented isolation paths, direct wl_pointer delivery and a separate wl_seat, and lays out an RFC-first plan covering permissions, threat model and upstreaming.
Catalog workOriginal source