Echoverse: Deep, Evolving Environments for Computer-Use Agents

Summary

Microsoft introduces Echoverse, a suite of twelve training worlds for computer-use agents: ten deep domain worlds and two capability worlds, each drilling a single control (e.g., date pickers, nested filters) in varied forms. Depth is key: these worlds replicate real application behavior, come seeded with realistic data, and maintain coherent state across screens and users. Training on all twelve nearly doubles a 9B model's base score from 36.5% to 67.1%, coming within fourteen points of GPT-5.4. Lessons include: high simulation fidelity is essential, as shallow worlds hurt performance; agents struggle with common UI elements, but varied drilling teaches generalization to unseen domains; co-evolving the model, world, and verifier improves all; and reinforcement learning with a grounded verifier as reward surpasses imitation, improving held-out performance and reducing steps to goal. Four worlds are released with code, data, and grounded graders.