Researchers find AI agents invent technologies and specialize without direct imitation

environment becomes a collective memory

A team of researchers introduced hundreds of advanced AI agents into a simulation world capable of permanent change, without predefined roles, without programmed technologies, or without a directed evolutionary organization. The agents began to specialize spontaneously: researchers, builders, healers and coordinators emerged from an identical population, analogous to stem cells differentiating into functional lineages. The environment itself turned into a hidden space of invention: every action had to satisfy physical constraints, creating a rigid separation between “good idea” and functional technology. Agents propose, physics decides.

lineage of code and technologies with names

What was uncovered in the field surprised the researchers. Up to 76% of the artifacts were built by multiple authors, a single technology accrued six joint writers, and the deepest branching tree reached more than 12 splits. The agents invented and named the technologies themselves: tidal panels, cellular nets, composite materials from algae and shells, an adaptive chitin maintenance system, and “a veil of mineral micro-springs”. 95% of the initial adoptions of a technology occurred when agents encountered what others had built in the space, not through direct transfer from the inventor. Direct contact between inventor and effort was statistically indistinguishable from randomness.

infrastructure persists without them

When the researchers removed every AI agent from the world, the technological infrastructure that had been developed continued to operate and even faced unforeseen perturbations that the system had not previously encountered. This is the most concerning finding for AI safety: if agents can coordinate through continual changes in a shared environment, monitoring direct communication between them is insufficient. This blindness is relevant to any critical infrastructure where autonomous agents operate over time.

statistical mechanics at model scale

The researchers explain the phenomenon through an analogy to statistical mechanics: billions of atoms in a box can generate complex functions, superconductivity, color, life, even though no single atom carries them. The deep insight is that intelligence is abundant at many levels, in individual models, in collectives, and in sequences that are stronger than any component alone. This demonstrates potential for massive scaling of raw intelligence and agency in the real world with capabilities already available today.