Source
Echoverse: Deep, evolving environments for computer-use agents
Microsoft created Echoverse, a set of twelve high-fidelity synthetic environments designed to train AI agents in computer use. The system prioritizes behavioral depth and co-evolution between the model and environment over sheer quantity of training data.
Why it matters
Since many critical tasks occur in private systems that are unsafe for live AI training, these synthetic worlds enable agents to learn complex workflows safely before being applied to real software like banking or health records.
The details
- A 9B model's score increased from 36.5% to 67.1% after training.
- Targeted training on date pickers and nested filters improved general UI navigation.
- Microsoft released four of the environments, including code and data, for research.
Show entities and relationshipsHide entities and relationships
In this article
Products
Topics
Companies
Organizations
Key connections
Microsoft built Echoverse for training computer-use agents
Microsoft created EchoMail as part of the Echoverse environment suite
Microsoft owns EchoCalendar
Microsoft created EchoCalendar as part of the Echoverse environment suite
Microsoft created EchoChat as part of the Echoverse environment suite
Microsoft created EchoML as part of the Echoverse environment suite
Microsoft created EchoForge as part of the Echoverse environment suite
Show 16 more connectionsShow fewer connections
Microsoft created EchoBank as part of the Echoverse environment suite
Microsoft created EchoCare as part of the Echoverse environment suite
Microsoft created EchoForum as part of the Echoverse environment suite
Microsoft created EchoTunes as part of the Echoverse environment suite
Microsoft created EchoStay as part of the Echoverse environment suite
OpenAI developed the GPT-4.1 AI model
OpenAI owns GPT-4.1 vision
OpenAI developed the GPT-4.1 vision model
Echoverse is built with FastAPI
Echoverse synthetic backend environments are built with FastAPI
Echoverse is built with SQLite
Echoverse synthetic backends use SQLite databases
Echoverse synthetic interfaces are built with React
Echoverse uses Supervised Fine-Tuning
Echoverse outputs SFT trajectories for model fine-tuning
Echoverse uses Reinforcement Learning
Echoverse functions as an RLE for reinforcement learning
EchoStay uses InsideAirbnb
EchoStay is seeded from InsideAirbnb listings, hosts, reviews, and amenities
Echoverse is related to Hugging Face
Echoverse datasets and code are released on Hugging Face
Echoverse uses Browserbase
Live web benchmarks were evaluated through Browserbase
Qwen3.5-9B competes with GPT-5.4
Qwen3.5-9B trained on Echoverse was benchmarked against frontier model GPT-5.4
Related events
Release of Echoverse Training Environments
Get the weekly recap
The stories like this one, picked and explained — once a week, straight to your inbox.