Microsoft released 12 synthetic environments to improve computer-use agents

Microsoft researchers introduced Echoverse, a collection of 12 synthetic environments designed to train agents for computer-based work. By simulating applications like email, banking, and travel booking systems, the researchers created a space where AI can perform actions and receive feedback based on actual database states rather than visual screenshots. This method allows the models to learn how their clicks and inputs change underlying data, which is essential for automating complex workflows that rely on coherent state across different screens and users.

Testing showed that a 9B parameter model improved its performance from 36.5% to 67.1% after training on these environments. The researchers found that depth is critical; when models trained on shallow, less realistic versions of these sites, they struggled to generalize. By using reinforcement learning against the ground truth of the database, the agents learned to complete multi-step goals more efficiently while reducing the total number of steps required.

The project specifically targets common UI bottlenecks like complex date pickers and nested filters by offering capability-focused worlds. These environments force the AI to handle variations of specific controls that it might not have encountered in standard training data. By co-evolving the model, the environment, and the verification tools, the researchers were able to create a feedback loop that continuously increases task difficulty as the agents become more proficient.

Echoverse is limited to the specific tasks and application logic provided in these synthetic environments. While these worlds are designed to mimic real-world software, they do not replicate every edge case or external dependency found on live production websites. The researchers noted that these tools are intended to help AI models bridge the gap between simple imitation and robust autonomous execution in closed enterprise systems.