Synthetic Computer Environments for Agent Training

Synthetic Computers at Scale proposes a data generation method that creates synthetic computers resembling real per-user work environments at scale, simulating long-horizon productivity tasks. Directory structures are organized to look realistic, and filled with content-rich artifacts like documents, spreadsheets, and presentations, so agents can read context and continue tasks.

The simulation runs on a division of labor between two agents.

• One agent creates a productivity goal tailored to a specific user. This goal requires multiple specialized deliverables and is designed to need roughly a month's worth of work by human standards.
• The other agent uses the computer in that user's role, exploring the filesystem for grounding, coordinating with simulated collaborators, and producing artifacts until the goal is complete.

In initial experiments, 1,000 synthetic computers were created, and each simulation logged 8+ hours of agent runtime with an average of 2,000+ turns. The experience signal obtained this way led to meaningful performance gains on both in-domain and out-of-domain productivity evaluations, and the authors believe that, since personas exist at scale, this approach could extend to millions or even billions of environments.

Ultimately, this leads to the claim that the method could serve as a foundation for agent self-improvement and agentic reinforcement learning on long-horizon productivity tasks.