Francesco Bonacci is the founder and chief executive of Cua, an open-source computer-use infrastructure company building environments and tools that let AI agents operate desktop applications. A former Microsoft engineer and coauthor of Windows Agent Arena, he focuses on making those agents reliable, observable, and capable of working without taking over a user’s screen.
Bonacci began with Windows Phone, Xamarin, .NET, and Azure development before joining Microsoft in 2019. His work there spanned Xbox and Microsoft AI, where he helped develop Windows Agent Arena, a benchmark testing multimodal agents against more than 150 tasks inside a complete Windows operating system. The research, published at ICML 2025, found that its strongest evaluated agent completed 19.5 percent of tasks, compared with 74.5 percent for humans.
After leaving Microsoft in early 2025, Bonacci developed Lume, an open-source tool for running macOS virtual machines on Apple Silicon, and founded Cua that March. He announced its launch publicly in April. The Cua project has since expanded into cloud desktops, isolated agent sandboxes, cross-platform automation, and evaluation infrastructure.
Realistic operating-system evaluation. Windows Agent Arena measures whether agents can complete actual desktop tasks, not merely produce convincing interface actions. Bonacci’s contribution helped establish full operating systems as a necessary testing ground for software that must navigate applications, settings, documents, and unpredictable interfaces.
Isolated execution through Lume. Native macOS virtual machines give agents controlled environments without requiring unrestricted access to a developer’s everyday computer. That virtualization foundation informed Cua’s broader sandbox and cloud-desktop infrastructure.
Background computer use with Cua Driver. Bonacci’s technical writing on macOS window internals describes combining accessibility trees, screenshots, and process-directed input so agents can interact with applications while the person retains foreground control. Accessibility actions come first; pixel-based interaction supplies an alternative when application interfaces expose insufficient structure. Support differs across macOS, Windows, and Linux, and some applications still require foreground activation.
Agent productivity constrained by human attention. Bonacci has described vibe coding paralysis, in which abundant AI assistance creates unfinished projects and fragmented supervision. His practical goal is automation that preserves human visibility, control, and capacity to finish work.
Running agents beside a human requires more than another cursor: window-level control, trustworthy task evaluation, and infrastructure that keeps GPUs from waiting on desktops.
▶ Watch ↗