Microsoft researchers built Echoverse, a system that generates stateful synthetic applications for training computer-use agents through a co-evolution loop in which every graded training run doubles as both a training signal and an instruction to improve the environment itself. Prioritizing environment depth and continuous repair over sheer quantity, the team trained a 9-billion-parameter model to 67.1% accuracy across fourteen evaluation splits — within fourteen points of a frontier model. The suite mixes ten full-domain worlds (login-gated workflows like email and banking) with capability worlds that systematically vary UI controls, and the authors show shallow synthetic environments actually degrade performance while deep ones lift held-out task accuracy from 58.8% to 68.0%.
