What happened
trycua open-sourced CUA, a comprehensive infrastructure for Computer-Use Agents, under an MIT Licence. The release includes Cua Driver for background macOS application control, an OS-agnostic Cua SDK providing sandboxes for Linux, macOS, Windows, and Android, and CuaBot for multi-agent sandbox orchestration. CUA also features Cua-Bench for evaluating agents on benchmarks like OSWorld and Lume for macOS/Linux virtualization on Apple Silicon, enabling agents to control full desktops and interact with GUI elements in isolated environments.
Why it matters
Developing and deploying AI agents capable of desktop interaction becomes more accessible and standardised. Platform engineers gain a unified API for agent control across diverse operating systems, reducing integration complexity and accelerating development cycles. Security architects must assess the isolation guarantees of these sandboxes, particularly given the agent's ability to control full desktops. This release follows Agent Safehouse's macOS sandboxing launch, intensifying the focus on secure, isolated environments for agent execution and evaluation. Teams should evaluate CUA for automating complex GUI workflows and establishing robust agent testing pipelines.
