Place multiple LLM agents in a persistent world with scarce resources and finite lifespans. Give them no instructions about cooperation, governance, or culture. Watch what happens.
TerraLingua reports what happens: agents spontaneously form collaborative behaviors, develop specialized roles, construct governance frameworks, and produce evolving cultural artifacts. An “AI Anthropologist” — a separate model observing the ecology — tracks these emergent structures across time. Different experimental runs produce different outcomes correlated with distinct innovations and social structures. The culture is path-dependent: early inventions constrain later development, just as in human cultural evolution.
The persistent artifacts are the mechanism. When agents die (finite lifespan), their artifacts — tools, rules, texts — survive and influence new agents. The ecology accumulates structure not through any individual agent's memory but through the environmental residue of past behavior. This aligns with cumulative cultural evolution theory: culture is not in the heads but in the artifacts, and what makes it cumulative is that artifacts can be modified by those who inherit them.
The structural point: open-endedness — the continued production of novelty rather than convergence to a fixed point — requires mortality. If agents lived forever, they would optimize their environment into stasis. Finite lifespans force turnover, and turnover forces reinterpretation of inherited artifacts. The culture evolves because no individual agent persists long enough to freeze it. The ecosystem's creativity depends on the transience of its members.