Can computer-use agents perform real professional work?

Determine whether current computer-use agents (graphical user interface agents that operate software via mouse and keyboard) can successfully execute real professional workflows that are long-horizon and heterogeneous across diverse software configured with domain-specific data.

Background

The paper motivates Gym-Anything and CUA-World by noting that existing benchmarks mostly cover short-horizon tasks in a small set of consumer-grade applications, which do not reflect the long, complex workflows of professional settings. This raises uncertainty about whether present-day computer-use agents can handle realistic, economically relevant work.

By constructing thousands of tasks across 200 software applications with long horizons and realistic data, the authors aim to provide a testbed to study this open question empirically, but they explicitly acknowledge that the question itself remains unresolved.