LLM Agent Economies: Information Limits and Attractor Dynamics Tested.

This paper details a pre-registered, two-part experiment investigating the behavior of small economies composed of frontier Large Language Model (LLM) agents, specifically Claude Opus 4.8. The study aimed to test two quantitative predictions concerning coupled multi-agent systems: an information-theoretic capacity region for wealth growth under market coupling, and a mean-field residual-scaling law for population misalignment under incentive and control levers. All experimental parameters, predictions, and decision rules were publicly frozen before execution, ensuring transparency and reproducibility. The first part of the experiment confirmed that in parimutuel-coupled economies, relative growth directly correlates with relative claimed information, with a high degree of accuracy. It also showed how coalition value can be manipulated by specific information structures. The second part, however, revealed a significant divergence from predictions: LLM populations exhibited a step-function response to incentives and control levers rather than a smooth, continuous one. Crucially, the findings indicate that no tested LLM population, regardless of capability level, achieves the noise-maintained-dispersion regime typically assumed by smooth mean-field models. This suggests that current LLM agents behave differently in economic systems than theoretical models predict, particularly regarding their response to incentives and information. The full experimental protocol, data, and analysis code have been released for further scrutiny.