RealCUA Dataset Captures 600 Hours of Computer-Use Work

Hunting for a dataset that shows how people actually get work done on a computer, rather than how they click through a tidy demo? RealCUA is aimed at exactly that gap. It is being introduced as the first computer-use dataset built from open-ended work in environments that keep changing.

The headline number is 600+ hours, and it is still growing. That is real multi-objective work, with tasks interleaved across thousands of steps on realistic desktops.

What Happened

RealCUA is described as the first computer-use dataset of open-ended work in evolving environments. It holds more than 600 hours of recorded work, a figure the source says is still growing.

The work behind it is multi-objective rather than scripted: people juggling more than one goal at a time, with tasks that interleave, spread across thousands of steps. The recordings come from realistic desktops, and the environments evolve rather than staying frozen from start to finish.

Why It Matters

If you build or test computer-use agents, the hard part has never been the single clean task. It is the switching, the half-finished work, and a screen that shifts under the agent's feet. A dataset that keeps those conditions instead of stripping them out is closer to what an agent meets on a real machine.

The 600-hour figure matters too. Thousands of interleaved steps give you enough recorded behavior to study, not just a handful of cherry-picked clips.