RecreationWorld: Scalable and Verifiable Environments for Hybrid Computer-Use Agents

RecreationWorld. Scalable and Verifiable Environments for Hybrid Computer-Use Agents. By Shuai Bai and thirty-one others, from Alibaba Token Hub, Alibaba Group. Published on the eighteenth of September. Its code repository is worked out: about two hundred source files, with a readme and a licence.

Computer-use agents have developed along two separate lines: graphical interaction, and software development through code and the command line. Real digital work needs both, interleaved rather than stacked end to end. The authors study hybrid agents that decide for themselves when to explore an interface, when to implement software, and when to run and visually verify what they have built. Measuring that requires a task in which none of those three can be skipped.

The organizing task is recreation. Given a running reference application, an agent has to discover how it behaves and build a faithful implementation, with no workflow prescribed. RecreationWorld offers this across five platforms: Ubuntu, macOS, Windows, Android, and the Web. It provides reproducible environments and a unified harness with native GUI control and coding tools. The running reference acts as an oracle for hidden behavioral tests, which gives rewards grounded in execution. Because the reference is running, it can be exercised the same way as the agent's version, and the two compared behavior for behavior.

For held-out evaluation there is RecreationBench, with two hundred and fifty diverse tasks across domains and platforms. Its assertions are both programmatic and visual, grounded in the reference, and they cover outcomes conditioned on actions at multiple interaction depths. Every assertion is validated on the reference and by human reviewers before the suite is frozen for automatic scoring. The benchmark therefore tests not only whether an application looks right on first load, but whether it responds correctly after the user has done something, and after several things in sequence. And since every assertion was checked against the reference by machine and by people, a failure points at the agent's build rather than at the test.

On the training side, trajectory generation is scaled using high-quality open- source applications. Models trained on those trajectories improve across five out-of-distribution benchmarks for coding and hybrid computer use, and they verify their rendered outputs more often. The authors take that as evidence of transfer beyond recreation itself.

On RecreationBench, GPT-6 Astra leads at fifty-eight point one percent overall, yet passes every programmatic test on just two point eight percent of tasks. Agents reproduce the static structure of an interface more reliably than its interactions and computed outputs. And the applications they generate remain smaller and more monolithic than their references. Taken together, these findings locate the difficulty. Copying what a screen shows is within reach. Copying what the software does, across interactions and computations, is where agents fall short.

Recreation works as a task because the reference is always there to answer against. Whatever the agent builds can be compared with the real thing, action by action, and on that comparison even the leading agent seldom gets everything right. The benchmark, environments, and test suites are released.