Researchers have released a new paper on Hugging Face describing RecreationWorld, a framework for training and evaluating hybrid computer-use agents (CUAs) — systems that combine graphical interface interaction with software development through code and the command line.
The paper argues that real digital work requires both modes of operation interleaved rather than stacked end to end. Hybrid CUAs must autonomously decide when to explore an interface, implement software, and run and visually verify their artifacts.
RecreationWorld is built around the task of recreation: given a running reference application, an agent must discover its behavior and build a faithful implementation with no prescribed workflow. The framework provides reproducible environments on Ubuntu, macOS, Windows, Android, and Web, plus a unified harness with native GUI control and coding tools. The running reference acts as an oracle for hidden behavioral tests, supplying execution-grounded rewards.
The authors scale trajectory generation using high-quality open-source applications. Models trained on these trajectories improve across five out-of-distribution coding and hybrid computer-use benchmarks and more frequently verify their rendered outputs, which the paper presents as evidence of transfer beyond recreation.
For held-out evaluation, the team introduces RecreationBench, comprising 250 diverse tasks across domains and platforms. Reference-grounded programmatic and visual assertions cover action-conditioned outcomes at multiple interaction depths, and each is validated on the reference and by human reviewers before the suite is frozen for automatic scoring.
According to the results, GPT-6 Astra leads at 58.1% overall, but passes all programmatic tests on just 2.8% of tasks. Agents reproduce static interface structure more reliably than interactions and computed outputs, and generated applications remain smaller and more monolithic than their references. The benchmark, environments, and test suites are being released.