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

Echoverse introduces a framework that compiles specs into stateful applications with grounded graders for training computer-use agents.

It uses a co-evolution loop to repair environments while training agents, making login-gated and other complex web tasks more learnable.

Across 12 environments, a 9B model improved from 36.5% to 67.1% over 14 evaluation splits.

The authors also released four environments as a benchmark, showing that environment quality and depth can matter as much as scale.