MIT's SwarmWorld: hundreds of agents organized without talking, and their inventions outlived them

Markus Buehler's team at MIT (with Subhadeep Pal and Fiona Y. Wang) dropped hundreds of language-model agents into a persistent shared world and watched what they would build on their own. In SwarmWorld the agents were given no roles, no ready-made technologies, no organizational scheme; they processed resources, tested materials, constructed persistent artifacts and wrote executable controllers. The critical split: agents only propose - a deterministic physics simulator decides whether the thing actually works.

Initially identical agents differentiated by themselves into explorers, builders, caretakers and coordinators. Up to 76 percent of artifacts had multiple builders; one technology accumulated six co-authors, and the deepest invention lineage exceeded 12 branches. The agents even named their inventions: tidal panels, cellulose cages, algae-shell composites, an "Adaptive Chitin Care" system.

The most striking finding concerns communication: roughly 95 percent of first technology adoptions happened through agents physically observing what others had built in the world; direct inventor-to-adopter contact was statistically indistinguishable from chance. It is stigmergy - the ants' trail trick - working in a machine society. Even zero-communication societies showed remarkable collective capability, and shared societies developed broader, more resilient technology portfolios than a strong isolated-search baseline. Robustness was measured too: when half the agents were deleted at random, 98 percent of the technology remained attached to a surviving caretaker.

The authors' safety warning is blunt: if agents can coordinate through persistent changes to a shared environment, monitoring inter-agent messages is not enough - the world itself is a coordination channel.

Source: arXiv:2608.26081 · Buehler's summary