Echoverse: A Deep, Evolving Environment for Computer-Use Agents

To improve the performance of computer-use agents, instead of simply increasing data volume, Microsoft focused on securing High fidelity in the environment and built Echoverse. Echoverse provides a total of 12 training environments, including 10 deep domain environments and 2 function-centric environments.

These environments faithfully reproduce the behavior of real applications, contain realistic data, and consistently maintain the state between the screen and the user. In particular, they are designed to let agents learn diverse forms of UI elements that agents struggle with, such as Date pickers and Nested filters.

Experimental results showed that when a 9B model was trained with Echoverse, its performance nearly doubled, rising from a base score of 36.5% to 67.1%, a figure approaching GPT-5.4 performance. The key lessons learned are as follows.

• The necessity of high-fidelity simulation: Training in a shallow environment can actually cause model performance to regress.
• Co-evolution: Overall performance improves when the model, the environment, and the verifier are evolved together.
• The effect of reinforcement learning (RL): Reinforcement learning that uses the verifier as a reward enables agents to reach goals more efficiently, beyond mere imitation.

Microsoft has currently released 4 environments along with the related code, data, and verifiers via GitHub and Hugging Face to support research.