jev-use replaces coordinate clicks with a list of screen regions

Cua @trycuaOriginally posted 2026.09.28

Open original on X

5/ In jev-use, our reference agent loop, regions become a closed list of candidate actions. A decision model such as Cua-S1-4B picks one ID, not coordinates. Real-time and text-only models were never trained to point at pixels. With regions, they can still act on the screen.