Poor Man's Agentic Modeling: Simulating Large LLM-Agent Societies on a Laptop

Details

Simulating societies of hundreds or thousands of LLM agents to study macroscopic behavior — phase transitions, stylized economic facts, scaling laws — is expensive. But most research questions in this area are about collective behavior, not individual agent cognition. Igor Itkin exploits a statistical physics insight: replace each LLM agent with a low-parameter surrogate model fitted from a small number of cheap queries, then run the society at any scale on a laptop. Whether surrogates work is determined before the simulation runs, using an interaction order × memory taxonomy that maps agent perception and memory structures to an effective theory and a predicted error trend. This taxonomy is validated on a faithful reimplementation of EconAgent (an LLM macroeconomy simulation) and seven other named LLM agent simulations, with decisions cloned from genuine DeepSeek LLM elicitations for a few dollars total. The predicted error trends hold cell by cell across the taxonomy. Two refuted predictions — both on strongly saturating responses traced to their curvature — are quantitatively explained by the theory with no free parameters. Code and data are publicly released. For AI researchers running agent-based modeling studies, this is a practical cost-reduction tool potentially reducing simulation expenses by orders of magnitude for macroscopic research questions.