收录解读
This paper proposes synthetic user computers as scalable environments for long-horizon productivity-agent training and evaluation, including realistic folders, artifacts, collaborator context, and multi-deliverable objectives.
The reported setup runs agents for thousands of turns over many hours in simulated professional computer environments, generating experiential learning signals from extended task execution rather than short benchmark prompts.
Its central contribution is an environment-generation pattern for agent self-improvement: create diverse synthetic workspaces, let agents work through month-scale objectives, and use the resulting traces to improve later agents.
For this repository, the paper is important as an agent-training substrate and evaluation direction for real computer-use productivity workflows.