MIT LLM Agents in Simulation Spontaneously Divided Into Roles

🤖 MIT LLM Agents in Simulation Spontaneously Divided Into Roles

In the arXiv:2608.26081 preprint "SwarmWorld", hundreds of identical LLM agents in a MIT simulation without roles or direct communication spontaneously divided into researchers, builders, artifact keepers, and coordinators. Coordination occurred through stigmergy — traces in a shared environment — and the societies produced broader and more robust solution portfolios than best-of-N sampling.

🌍 The work offers the industry an alternative to orchestrating multi-agent systems through chat: coordination happens through a shared environment, and function is separated from the model's "opinion" and verified by a simulator after the agents are removed.

👤 The "researcher/builder/keeper/coordinator" specialization emerged without role prompts — from the environment's structure and simulator selection. A guide for practitioners: where cooperation expands the solution set, and where best-of-N is sufficient.

Source 1: https://arxiv.org/abs/2608.26081 Source 2: https://arxiv.org/html/2608.26081v1