Markus Buehler's group at MIT posted a paper that is both a finding and a caution. They put a hundred identical language-model agents into a simulated world the agents could permanently change, with no assigned roles, no predefined technologies, and no built-in organization. The agents specialized on their own, built persistent artifacts, inherited and modified each other's executable code, and turned the environment into a record of what the society had learned. When the researchers removed every agent, the technology the agents had built kept operating and held up under disturbances none of them had seen.
The paper is SwarmWorld: Stigmergic technological evolution in societies of language-model agents, by Subhadeep Pal, Fiona Y. Wang, and Markus Buehler at MIT's Laboratory for Atomistic and Molecular Mechanics. Buehler's post announcing it is stronger than the paper itself.
The setup: agents propose, physics decides
SwarmWorld is a spatial simulation. Agents explore, process resources, test materials, construct artifacts, and write executable controllers. The key design move is that the agents propose, and a deterministic simulator decides. An idea only becomes a technology if it clears a physical validation gate, which puts a hard line between a plausible claim and something that actually works. The paper's phrase for this is that SwarmWorld "splits cognition from consequence."
The experiment runs four conditions. Full culture has a shared world plus direct messages, records, and cross-agent program inheritance. No explicit culture keeps the shared world but strips the explicit channels. No communication lets agents coordinate only through the world itself. Independent search puts each agent alone in its own copy of the world.
What emerged
From identical starting agents, the societies differentiated on their own. Looking back at the behavioral data, the authors found agents had sorted into exploration, construction, maintenance, and coordination roles. In the full version, a little over half the agents ended up focused on the shared artifacts.
The technologies developed family trees. Agents copied and changed code others had written. Many artifacts had more than one builder, code lineages averaged about ten forks deep under full culture, and the agents even named their own inventions: tidal panels, cellulose trellises, kelp-shell composites, an "Adaptive Chitin Maintenance" system, a "Mycelial Mineral Spring Veil."
The diffusion result is the one worth sitting with. Roughly 95% of first technology adoption happened through physically observing an artifact in the world, not through a direct handoff from the inventor. Direct inventor-to-adopter contact showed no consistent edge over shuffled contact. The agents mostly learned a technology by walking past it. That is stigmergy, the coordination-through-environment trick that termites use.
The animation below sketches that loop. The dots are the agents. They interact only with the world, not with each other. One builds something, another walks by and builds on what it found. It is a stylized drawing of the mechanism, not a frame from the paper's simulation.
Then the knockout test. Removing half the agents at random left almost all the technology connected to a surviving caretaker. Removing the high-traffic hub agents was much worse. Buehler's post puts the collapse at about 60%, while the paper itself only notes "targeted vulnerabilities."
The fine print
One caution before the editorial. The post frames this as non-communicating societies winning. The paper is more careful. Shared worlds build a wider, tougher set of tools than lone agents, but the single best tool is a closer contest that moves over time, and the no-explicit-culture version sometimes beats full culture.
Editorial
The world as memory
The part of this I actually like is stigmergy. It is a simple idea. Instead of talking, you leave marks on the world, and the next agent reads those marks and builds on them. Termites build huge nests this way without any termite having a plan. What this paper tests is whether AI agents can do the same.
The part I keep thinking about is memory. The agents stored their work in the world itself. When the researchers deleted the agents, the work was still there, in the physical things they had built. That is intelligence living in the arrangement of objects, more like a coral reef or a city than like a language model.
Two things to keep in mind. First, this happened inside a simulation with a simple physics engine. It is a strong sign, not proof about the real world. Second, the safety worry cuts both ways. Buehler says it is dangerous because you cannot catch agents coordinating if they never talk. That is true. The flip side is that things built out in the open are easier to inspect than things said in private. I see that as a good thing.
Amber diamonds are artifacts. When an agent reads one, a ring expands from it and the agent builds a related artifact nearby, leaving a faint dashed line to its parent. The message counter stays at zero the whole time. The artifacts still accumulate.