RecreationWorld is a five-platform (Ubuntu, macOS, Windows, Android, Web) framework that trains hybrid computer-use agents to both explore interfaces and write code by tasking them with faithfully recreating a running reference application, using the reference itself as an oracle for hidden behavioral tests. Trajectories generated against real open-source applications improve trained models across five out-of-distribution coding and computer-use benchmarks. The accompanying RecreationBench, with 250 human-validated tasks, shows even the strongest evaluated model (GPT-6 Astra, 58.1% overall) passes all programmatic checks on only 2.8% of tasks, exposing a gap between surface GUI competence and reliable functional implementation.
