StateAct: Program State, before Pixels, for Long-Horizon Computer-Use Agents

Researchers introduced StateAct, a computer-use agent that reasons primarily over program state instead of screenshots.

A visual specialist is used for only a small share of steps, reducing token usage and cost while improving long-horizon task handling.

StateAct set new top results on OSWorld 2.0 and improved performance on OSWorld-Verified, WindowsAgentArena, AndroidWorld, and MobileWorld.

The approach also delivered strong gains with the 31B SFR-CUA model and on Claude Opus 4.8, suggesting a new direction for UI automation.