Agent World Model in the Wild: Computer-Use Agent Environment Synthesis

Agent World Model in the Wild. Computer-Use Agent Environment Synthesis. By Zhaoyang Wang and eleven others, from the University of North Carolina at Chapel Hill, Microsoft Research, and three other institutions. Published on the nineteenth of October. Its code repository is worked out: about nine hundred source files, with a readme.

Agent learning is moving away from learning from human data and toward learning by interacting with the world. Scaling that kind of learning directly on real computer systems is difficult, and it can introduce substantial security risks: an agent practising on a real machine can do real damage. Synthetic environments are the alternative, but existing approaches typically specialize in a single way for an agent to interact.

Agent World Model in the Wild, or WildAWM, is a scalable pipeline for synthesizing controllable environments and tasks to train general computer-use agents, or CUAs. The environments it builds are executable, and each one supports three interfaces together: graphical user interfaces, model context protocol services, and command-line interfaces. An agent inside one can therefore combine visual interaction, tool use and code execution as the task demands. The whole pipeline is powered by open-source models, which the authors take as a demonstration that high-quality computer-use environments can be generated without proprietary frontier models.

Environments are only half the job. Tasks have to be synthesized too, and the authors address three challenges that recur there. A task must be solvable. It must stay difficult enough to teach something. And its verifier, the check that decides success, must be reliable.

The answer is a set of quality gates based on execution. They combine rollouts from heterogeneous agents, consistency checks across different judges, and counterfactual probes that test whether a judgment is robust. A task passes into the training set only after surviving these gates.

The resulting pipeline produces three hundred and forty-one long-horizon pairings of environment and task. Some can take more than one hundred agent steps to solve, and they remain challenging for frontier models: GPT-6-Astra solves only sixty percent of them.

Two Qwen models of different sizes are then trained exclusively inside these synthetic worlds. Both achieve substantial gains across eight out-of-distribution benchmarks, spanning web agents, claw agents and computer-use agents. The authors take this as evidence that learning in WildAWM's synthesized worlds generalizes broadly to computer-use tasks the models have not seen. The code is publicly available.

Two choices define the pipeline. One is breadth of interface, so that an agent trained in a synthetic world can click, call tools and run code within a single task. The other is strictness about tasks, so that what enters training is solvable, hard and reliably judged. Built that way, and with open models only, the synthetic worlds were enough on their own to train agents that improved on benchmarks outside them.