Computer-use agents learn from the consequences of their actions, so training them requires applications they can interact with, break, and reset. Real-world applications are often login-gated and stateful, making synthetic environments a practical stand-in. Recent pipelines generate such environments in bulk, shifting the bottleneck from quantity to the quality of what lies inside each one.
Researchers at Microsoft Research identify three properties that drive returns: behavioral depth, targeting the interactions where agents actually fail, and the ability to improve alongside the model. They present Echoverse, which compiles specifications into stateful applications whose tasks are graded against the application's own database. A co-evolution loop reads every graded rollout twice: once to repair the environment, its tasks, and its verifier, and once to provide training signal for the model.
Trained on twelve such environments, a 9B model improves from 36.5% to 67.1% across fourteen evaluation splits, coming within fourteen points of the much larger frontier model that taught it. The paper examines each property in turn: shallow environments can actually hurt live-site accuracy (dropping from 80.0 to 75.0), while deep ones raise it (from 80.0 to 85.0 and from 48.0 to 65.0). Drilling one interface control across many renderings transfers to held-out widget families and the open web. Repairing a single environment lifts a model trained on it from 16.2% to 38.5%.
The same worlds serve as reinforcement-learning environments. A reward combining the grounded verifier with a dense per-step judge raises held-out score from 58.8% to 68.0%. The authors release four environments as a benchmark, complete with applications, seed data, and grounded graders. Code and additional resources are available at the provided links.