Cua + Scale AI Launch Cua Driver Bench: Background Desktop Agents vs Undisturbed Users; Vendor macOS Full-Pass ~58% for cua-driver

Core Background & Industry Pain Points

Computer-use agents often finish desktop work by stealing focus, moving the pointer, and raising windows—fine on a spare machine, unsafe on the laptop you are typing on. On 2026-11-06, Cua and Scale AI AgentEnv launched Cua Driver Bench to test whether a driver can finish real desktop tasks in the background without disturbing the foreground user or cheating through a shell.

Architecture Highlights & Internals

Runs pin Claude Sonnet 5.5, prompts, and limits (360s / 45 turns) and swap only the computer-use driver (cua-driver, Vercel agent-desktop, browser-use, Codex Computer Use). Ten public Bench v2 tasks cover multi-app workflows, precision interactions, and interruption dialogs. In the background condition a User activity window keeps typing; a full pass needs task done, GUI-only, and none of eight disturbance criteria. Cua discloses it builds cua-driver and prefers Scale numbers where available.

Authoritative Benchmarks & Measured Scores

Figures below are vendor/Scale-reported, not independently reproduced here. macOS background full-pass: cua-driver 23/40 (~58%), agent-desktop 8/40 (~20%), browser-use 1/40 (~3%); Codex Computer Use (Cua-run) 13/40 (~33%), overlapping cua-driver at the edge. macOS foreground: cua-driver 39/50 (~78%). Windows (Cua-run) cua-driver 11/40 (~28%); Ubuntu XFCE about 13%, with modal dialogs often taking the keyboard.

Developer Hands-on Guide

Read GitHub blog/cua-driver-bench.md and libs/cua-driver-bench; Scale trajectories live on HF Space/dataset earakely-scale/CuaDriverBench-AgentEnv. Prefer background full-pass plus disturbance splits over foreground completion alone; hover and background drag still depend on OS primitives.