RecreationWorld is a five-platform framework (Ubuntu, macOS, Windows, Android, Web) where hybrid agents autonomously recreate a running reference application by interleaving GUI exploration, coding, and . The framework provides reproducible environments, a unified harness, and hidden behavioral tests grounded in the reference. Trajectories are generated using high-quality open-source applications, and models trained on these trajectories improve across five out-of-distribution coding and hybrid benchmarks, while also verifying rendered outputs more frequently.
Evaluation on RecreationBench, a suite of 250 diverse tasks, shows that GPT-6 Astra leads at 58.1% overall, but passes all programmatic tests on only 2.8% of tasks. The study highlights that agents reproduce static interface structure better than interactions or computed outputs, and generated apps are smaller and more monolithic than references. This work is interesting because it enables grounded, reproducible training for hybrid computer-use agents and demonstrates transfer to unseen tasks.