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

Researchers introduced Echoverse, a framework for generating synthetic, stateful environments to better train computer-use agents.

Echoverse compiles specifications into login-gated-like applications and co-evolves environments, tasks, verifiers, and training signals from graded rollouts.

In experiments across twelve environments, a 9B model improved substantially on evaluation splits, and four environments were released as a benchmark.

The work suggests that deeper, task-relevant, evolving environments can materially improve agent performance and support reusable research benchmarks.