OfficeTown simulates organizations to improve AI agent teamwork and learning
OfficeTown: Generative Organizations for Evolving Agent Intelligence
Artificial Intelligence
Summary
AI agents often need to do work like people in professional settings, using files and tools while sharing information over time. The authors present OfficeTown, a virtual environment that creates simulated companies where AI agents take on roles, assign tasks, and check each other's work. This setup helps collect lots of meaningful work examples and feedback to train AI models better. Their tests show AI can improve by learning in these simulated work environments compared to simpler setups.
What this means in practice
- •For enterprise ai developers: Develop AI assistants that better collaborate on ongoing professional projects by simulating organizational workflows and task handoffs.
- •For software team leads: Use simulated organizations to test and improve AI tools for coordinated work and quality review before deployment in real teams.
Authors
Hongchen Wei, Daohong Ying, Zhi Hong, Aidan Wu, Bei Liu, Ruichun Ma, Yifan Yang, Qi Dai, Kai Qiu, Rui Xu, Dongdong Chen, Zhenzhong Chen, Chong Luo
Abstract
AI agents increasingly use files and software tools to produce professional deliverables. Such work draws on materials, decisions, and feedback accumulated across assignments. We introduce OfficeTown, a generative simulation framework for professional organizations in which tasks and their context arise from continuing organizational work. From a natural-language description of an organization's purpose and activities, OfficeTown initializes an organization whose members coordinate work, delegate bounded assignments to assistant agents, and review their deliverables. Artifacts, decisions, and unresolved matters persist and shape subsequent work, while responsibilities and permissions determine what each member can see and do. We curate work requests, their context, and review feedback into executable resources for evaluation and learning. We construct OfficeTown Bench with 406 tasks and 4,836 evaluation criteria from 23 simulated organizations across eight professional domains, incorporating domain-expert feedback. Frontier-model evaluations reveal substantial remaining headroom. Supervised fine-tuning and on-policy distillation on OfficeTown data substantially improve Qwen3.5-9B on external professional-work benchmarks, including GDPval. Supplementary evaluations show that the principal--assistant harness achieves higher scores than the single- and multi-agent baselines tested. These results highlight the value of organizational simulation in generating work-grounded data and environments for the continued development of generalist agents.