Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale

As computer-use agents move from simple information retrieval to complex, stateful tasks (like banking or cloud management), they require training environments that allow for safe, repeatable, and verifiable interactions. The authors investigate what makes a synthetic environment effective for training, specifically testing whether the quantity of environments or their internal "depth"—the fidelity of state, workflow dependencies, and grounded verification—is the primary driver of agent performance.

The authors introduce Echoverse, a pipeline that compiles specifications into stateful, login-gated applications. Unlike previous benchmarks that rely on visual appearance or model-based judgment, Echoverse environments are "grounded": success is determined by querying the application's own database after an action is performed. The researchers implement a co-evolution loop where every training rollout is used for two purposes: as a training signal for the agent and as a diagnostic tool to repair the environment, its tasks, and its verifiers. This ensures that the training data remains high-quality and challenging as the model improves.

The study identifies three critical levers for training: behavioral depth, capability targeting, and co-evolution. The authors report that "shallow" environments—those that look realistic but lack deep state coherence—can actually degrade performance compared to a base model. In contrast, deep environments significantly boost accuracy. By training a 9B model on twelve such environments, the authors achieved a 67.1% success rate, closing the gap with larger frontier models. Furthermore, they demonstrate that these environments serve as effective reinforcement learning (RL) training grounds, where a grounded reward signal significantly outperforms standard imitation learning.