OfficeTown: Generative Organizations for Evolving Agent Intelligence

OfficeTown. Generative Organizations for Evolving Agent Intelligence. By Hongchen Wei and twelve others, from Wuhan University, the University of California, San Diego, and four other institutions. Published on the sixteenth of October. No code repository is linked to it.

AI agents increasingly use files and software tools to turn out professional deliverables. Real professional work, though, does not arrive as a stack of isolated requests. It draws on materials, decisions and feedback that have piled up across earlier assignments. A task handed to an agent with none of that history is a thinner version of the work it is meant to stand in for.

OfficeTown is a generative simulation framework for professional organizations, built so that tasks and their context arise from continuing organizational work rather than being written one at a time. You describe an organization's purpose and activities in natural language, and OfficeTown initializes an organization from that description. Its members coordinate work, delegate bounded assignments to assistant agents, and review what those assistants deliver.

Two properties make the simulation feel like an office rather than a queue. Artifacts, decisions and unresolved matters persist, and they shape whatever work comes next. And responsibilities and permissions determine what each member can see and do. Out of this, the authors curate work requests, their context, and the review feedback into executable resources that serve both evaluation and learning. One of those resources is OfficeTown Bench: four hundred and six tasks and four thousand eight hundred and thirty-six evaluation criteria, drawn from twenty-three simulated organizations across eight professional domains, with domain-expert feedback folded in.

Frontier models evaluated on the benchmark leave substantial headroom. On the training side, supervised fine-tuning and on-policy distillation on OfficeTown data substantially improve Qwen3.5-9B on external professional-work benchmarks, GDPval among them. That is the result to weigh most, since the gains show up outside the simulation that produced the data. Supplementary evaluations add that the principal and assistant harness scores higher than the single-agent and multi-agent baselines that were tested.

Underneath the benchmark sits a claim about where agent data should come from. Instead of collecting tasks, OfficeTown simulates the organization that would generate them, with memory and permissions intact, and lets the work accumulate. The authors read their results as showing the value of that kind of organizational simulation for producing work-grounded data and environments for generalist agents.