LLM agents spontaneously evolve culture with no top-down engineering

In a new preprint accepted at Alife 2026, Simon Jones and Sabine Hauert explore what happens when LLM agents operate with no context beyond a single turn, minimal prompting, and only basic tools. Inspired by swarm engineering, they created collectives of three agents capable of sending messages and manipulating a shared text store that actively decays over time. By introducing evolutionary pressure, they observed that the agents spontaneously developed cooperative behaviors, storage management strategies, and generated complex evolving cultural artifacts with no explicit instructions from above.

Using dynamical systems analysis, the team demonstrated that these emergent behaviors exhibit structured long-range coherence that extends far beyond the entropy horizon of the decaying store. This coherence is consistent with Sperber's definition of emergent culture—where cultural artifacts arise from local interactions without central control. The findings suggest that even minimal LLM systems can self-organize into cultural-like dynamics, opening new avenues for studying artificial societies and decentralized AI coordination.