Can Electric Sheep Evolve? Engineering Evolution in Populations of Language-Model Agents

The idea of artificial intelligence that evolves has long been a theme in science fiction. It also has roots in early research on complex adaptive systems and artificial societies, including Holland’s Echo framework and Epstein and Axtell’s Sugarscape, which modeled adaptation, resource competition, reproduction, and cultural transmission among artificial agents. Contemporary LLM development, however, has largely centered on pretrained models and gradient-based post-training, while explicitly Darwinian population processes have remained comparatively peripheral. This seminar reviews an emerging literature on evolutionary optimization, cultural transmission, and population dynamics in LLM-based agents. I distinguish systems that merely borrow evolutionary terminology from those implementing meaningful variation, heredity, and reproduction. I then assess the strengths and limitations of current studies and ask whether such processes can support cumulative adaptation, collective intelligence, and potentially new levels of organization among artificial agents.