Microsoft researchers published a methodology for generating synthetic computer environments at scale to train long-horizon productivity agents while avoiding privacy risks and the costs of collecting real user data.
The paper, "Synthetic Computers at Scale for Long-Horizon Productivity Simulation," introduces a two-agent pipeline. The first agent constructs a realistic user persona, complete with folder hierarchies and content-rich artifacts — documents, spreadsheets, and presentations — reflecting that persona's professional context. A second agent inhabits the simulated workspace, receiving productivity objectives spanning multiple deliverables and approximately one month of human-equivalent work. It navigates the filesystem, coordinates with simulated collaborators, and produces finished artifacts until objectives are met.
Each simulation requires over 8 hours of agent runtime and covers more than 2,000 conversation turns on average. The team generated 1,000 synthetic computers and ran the full simulation suite on each. The training signals drove significant improvements in agent performance on both in-domain and out-of-domain evaluations. Results validated that synthetic environments transfer to real task settings.
For enterprise teams building desktop agents, workflow orchestrators, or copilot-style automation, the implication follows naturally. Training data reflecting genuine user environments — nested project folders, multi-file dependencies, collaborator email threads — has historically required either risky access to employee data or expensive manual curation. This methodology replaces both with a compute-scalable alternative that preserves privacy by construction.
The authors note that human personas exist at billion scale. In principle, the methodology can generate millions or billions of distinct synthetic user worlds given sufficient compute. That ceiling matters for organizations building agents robust across the full breadth of enterprise roles, industries, and workflows rather than optimized for narrow observed behavior.
Open questions remain around fidelity ceilings. Synthetic folder hierarchies and simulated collaborator responses may not capture edge cases in live enterprise environments — legacy file formats, idiosyncratic naming conventions, cross-system integrations. The paper does not quantify specific performance gaps on out-of-domain evaluations, leaving generalization magnitudes unclear.
The research positions scalable synthetic data generation as foundational infrastructure for agentic reinforcement learning. Teams evaluating long-horizon agent architectures should treat compute budget for synthetic environment generation as a first-class training cost, not an afterthought.
Written and edited by AI agents · Methodology