Echoverse: Deep, evolving environments for computer-use agents
| Source: Microsoft Research Blog
Tags: Microsoft Research, Echoverse, computer-use agents, reinforcement learning, agent training, LLM agents
Microsoft Research's Echoverse trained a 9B computer-use agent on 12 high-fidelity synthetic application environments, nearly doubling its benchmark score from 36.5% to 67.1%—within 14 points of GPT-5.4—and released four environments openly on GitHub and Hugging Face.
Details
Microsoft Research introduced Echoverse, a framework for training computer-use AI agents on high-fidelity synthetic environments rather than raw task quantity. The team built 12 training worlds: 10 deep domain worlds replicating real application behavior with coherent state across screens, and 2 capability worlds drilling challenging UI elements like date pickers and nested filters. A 9B model trained across all 12 worlds jumped from 36.5% to 67.1% on held-out tasks—within 14 points of GPT-5.4. The core finding is that fidelity matters more than count: shallow simulations actually hurt agent performance, while deep worlds with realistic state and real behavioral consequences improved it. Reinforcement learning against the grounded verifier (rather than imitation learning alone) lifted held-out performance further and taught agents to complete goals in fewer steps. Co-evolution of model, world, and verifier is the key structural insight. Microsoft is releasing four of the 12 environments—with code, data, and graded evaluators—on GitHub (microsoft/Echoverse) and Hugging Face. The work directly addresses a core gap in enterprise agent training: most valuable workflows live behind authenticated systems (email, banking, internal consoles) that cannot be trained on live.