The useful question is what changes for users, developers or buyers, and whether the announcement stays industry context or becomes something people can actually use.

Microsoft’s research team built twelve training environments—ten deep domain worlds and two capability-focused worlds—to teach computer-use agents how to navigate real applications. Each world reproduces an application’s behavior with realistic data and coherent state across screens and users. Training a 9B model on these environments improved its base score from 36.5% to 67.1%, approaching the performance of GPT-5.4. The environments are designed to reflect real consequences of actions, such as state changes or message deliveries, which are essential for learning but absent in static screenshots or open, uncontrolled websites.

The team highlights that most consequential workflows—such as email, banking, or health records—reside in closed systems inaccessible for training. Synthetic worlds replicate these systems with controlled databases, allowing safe experimentation while preserving state and workflow dependencies. These worlds consist of an environment, tasks, and verifiers that grade outcomes against ground truth. The research emphasizes that depth—faithful reproduction of causal structures—matters more than sheer quantity of environments. A loop that iteratively improves both the environment and the model, rather than treating them as separate stages, yields compounding benefits over static benchmarks.

To create effective training data, the team generates tasks grounded in real database entities and verifies them against live applications. Each task undergoes rigorous checks to ensure plausibility and feasibility, with failures tagged to specific layers (e.g., database, frontend) for targeted fixes. The resulting tasks are used to fine-tune models like GPT-5.4, producing supervised fine-tuning data that reflects real-world constraints. The process is iterative: harder tasks expose gaps in the environment, which are then strengthened, creating a virtuous cycle where both the world and the model improve together.

The research demonstrates the importance of depth in training environments through experiments comparing shallow and deep worlds. On domains like Allrecipes and Hugging Face, models trained on deep worlds—where tasks span dependent steps—outperformed those trained on shallow worlds, which only rehearsed isolated actions. The team also developed specialized worlds for controls like date pickers and nested filters, which are ubiquitous but challenging for agents. Training on these worlds improved performance not only on the specific controls but also on broader tasks, indicating that agents learned generalizable rules rather than memorizing layouts.