Echoverse: Deep, evolving environments for computer-use agents

Microsoft Research describes twelve training worlds—ten domain clones and two capability worlds for date pickers and nested filters—with state that can be graded from a database rather than a screenshot. A 9B model trained on all twelve reportedly rose from 36.5% to 67.1%, within fourteen points of GPT-5.4.

The team says shallow versions of the same sites hurt or stalled performance, while deep worlds transferred; capability worlds lifted held-out widgets and modestly improved live-web scores. Four worlds, code, data and graders are being released.

This is Microsoft Research’s own blog, so scores, world counts and transfer claims are vendor-reported. The methodological claim is still the news: training against login-gated, stateful systems is safer in a synthetic world whose database can be reset and inspected.

Signal Desk's read: the depth-versus-count finding is the piece that should travel. Agents fail on controls and hidden state, not on missing screenshots. Building another shallow clone can teach the wrong reflexes.

The research loop—repairing the world, tasks and verifier from the same graded run—is more ambitious than ordinary fine-tuning. Its cost is that the factory itself becomes part of the product.

A world is environment, tasks and verifier. Domains include mail, calendar, chat, code, banking, health records, forums, media and travel. Evaluation on WebVoyager and Online-Mind2Web is presented as transfer, not training.