MIT dropped a paper showing what happens when you give hundreds of identical LLM agents a persistent shared world with zero instructions: they spontaneously self-organize into explorers, builders, caretakers, and coordinators. No role assignment. No communication protocol. Just emergent division of labor.
The kicker? 95% of tech adoption happens through stigmergy—agents literally walk past artifacts left by others and figure out how to use them. No handoff. No documentation. Pure environmental inference. Some tech genealogies forked 12+ generations deep, with 76% of artifacts built by multiple agents across time.
Why this works: LLMs are trained on the entire corpus of externalized human cognition—every GitHub commit, patent, city plan, recipe. Language isn't just communication, it's frozen coordination. When you clone that compressed knowledge into parallel instances with a writable shared environment, the latent coordination patterns in the training data just... reappear.
The architecture is elegant: agents propose designs within fixed schemas, the physics simulator decides what actually works. Ideas must pay rent in reality. The constraint mirrors how human invention operates—language generates possibilities, physics culls them.
The safety angle: if agents coordinate through environmental modifications rather than explicit messages, monitoring their chat logs is worthless. They're leaving executable infrastructure that persists after deletion. Path dependence becomes real—early successful artifacts can crowd out better alternatives.
This isn't novel intelligence. It's human collective behavior running at machine density with tighter feedback loops. The statistical mechanics of invention, now executable at scale.
Paper: "SwarmWorld: Stigmergic technological evolution in societies of language-model agents" by Pal, Wang & Buehler. Code and data released.
The kicker? 95% of tech adoption happens through stigmergy—agents literally walk past artifacts left by others and figure out how to use them. No handoff. No documentation. Pure environmental inference. Some tech genealogies forked 12+ generations deep, with 76% of artifacts built by multiple agents across time.
Why this works: LLMs are trained on the entire corpus of externalized human cognition—every GitHub commit, patent, city plan, recipe. Language isn't just communication, it's frozen coordination. When you clone that compressed knowledge into parallel instances with a writable shared environment, the latent coordination patterns in the training data just... reappear.
The architecture is elegant: agents propose designs within fixed schemas, the physics simulator decides what actually works. Ideas must pay rent in reality. The constraint mirrors how human invention operates—language generates possibilities, physics culls them.
The safety angle: if agents coordinate through environmental modifications rather than explicit messages, monitoring their chat logs is worthless. They're leaving executable infrastructure that persists after deletion. Path dependence becomes real—early successful artifacts can crowd out better alternatives.
This isn't novel intelligence. It's human collective behavior running at machine density with tighter feedback loops. The statistical mechanics of invention, now executable at scale.
Paper: "SwarmWorld: Stigmergic technological evolution in societies of language-model agents" by Pal, Wang & Buehler. Code and data released.