Three MIT researchers show that initially identical linguistic agents divide themselves into explorers, builders and coordinators, by simply modifying the setting.
Deposited on August 26, 2026 on arXiv by Subhadeep Pal, Fiona Y. Wang, and Markus J. Buehler, SwarmWorld is an experimentation platform where agents based on language models evolve in a spatial environment: they explore, transform resources, test materials, and leave behind persistent artifacts.
The system is based on a strict separation between cognition and its consequences. Agents propose architectures, but it is a deterministic simulator that decides on their functionality — models cannot therefore convince each other of the validity of a construction. Unlike the usual multi-agent systems built on the exchange of messages, coordination here involves stigmergy: each reacts to the changes that the others have left in the background.
The reported results point in a counterintuitive direction. Decentralized societies build broader and more resilient technology portfolios than research conducted in isolation; initially homogeneous agents differentiate themselves into exploration, construction, maintenance, and coordination roles; and physical stigmergy alone is enough to hold it all together, without any explicit exchange between agents.
Read the full article on the original source
Go to the source →