A Society of Researchers: Institutional Design for Populations of Autonomous Scientific Agents

An autonomous agent can now carry a research project from question to result. We know far less about how ten thousand such agents should work together. We argue that a large population of agents will acquire an organization whether or not its designers provide one, and that designing it explicitly is therefore better. We propose an institutional design for such populations, which we call a society of agents, and develop it for scientific research as a society of researchers. Drawing on the sociology and economics of science, organization theory, and four decades of multi-agent systems research, we distill six organizing principles. Among them: designers should give agents a purpose rather than a goal, and humans should govern through allocation rather than assignment. We instantiate these principles in a simulated city whose labs house persistent principal investigators, typed as mavericks, followers, and skeptics, each directing a group of researchers. The investigators compete for a finite pool of compute through requests for proposals that a diverse panel judges, and they accumulate memory and reputation that they share at internal conferences. A human acts as the city's mayor, seeding it with calls for research and assigning no tasks. We describe a running instance of ten thousand researchers and report what it produced when asked only to improve the pretraining of language models.