Poor Man's Agentic Modeling: Simulating Large LLM-Agent Societies on a Laptop
06:00 · August 13, 2026 · arXiv cs.AI RSS
Simulating societies of many large language model (LLM) agents is expensive, yet the questions asked of such simulations are usually macroscopic: phase behaviour, stylised facts, and scaling with the number of agents $N$, not the cognition of any single agent. We turn a statistical-physics observation into a method: replace each LLM agent by a low-parameter model fitted from a few hundred to a few thousand cheap queries, then run the society at any $N$ on a laptop. Whether this works is decided before the simulation runs, chiefly by what each agent perceives. We introduce an [interaction order x memory] taxonomy that maps perception and memory to an effective theory and a predicted $N$-trend of the surrogate error. We validate it on a faithful reimplementation of the LLM macroeconomy EconAgent and seven further named LLM simulations, with agent decisions cloned from genuine LLM elicitations (primarily DeepSeek) for a few dollars; the predicted error trends hold cell by cell, and the two refuted predictions, both on a strongly saturating response and traced to its curvature, are themselves matched quantitatively by the theory with no free parameters.
Summary
Simulating societies of large language model agents has remained costly because each agent typically invokes an expensive model at every step, even when the scientific questions concern only aggregate behaviour such as phase transitions, stylised facts, or scaling with population size N. The method presented replaces every LLM agent with a low-parameter surrogate whose handful of coefficients are fitted from a few hundred to a few thousand inexpensive queries, after which the entire society can be executed on a laptop for arbitrary N.
The decisive factor is the perception design of the original simulation. An interaction-order-by-memory taxonomy classifies each setting according to whether agents respond to a single population-wide statistic, to a community-level signal, or to local graph neighbours, and whether they retain memory of past states. Each cell of the taxonomy maps directly onto an effective theory and supplies a parameter-free prediction for how the surrogate error scales with N: vanishing as N to the minus one-half in the mean-field regime, remaining order-one when the input is heterogeneous, or growing when local structure is discarded by averaging.
The taxonomy was tested on a faithful reimplementation of the EconAgent macroeconomy together with seven further named LLM simulations whose agent decisions were cloned from genuine model elicitations, primarily DeepSeek. In every pre-registered cell the observed error trend matched the prediction; the two cases in which the scalar surrogate failed were both traced to strongly saturating response functions whose curvature was recovered quantitatively by the same theory. The approach therefore supplies a practical criterion, decided before any large run, for when macroscopic observables can be recovered without keeping the original language models in the loop.
Why it matters
This research is highly relevant for Dutch AI researchers and SMEs, offering a mathematically rigorous and computationally cheap way to simulate and study multi-agent systems. It aligns with the Netherlands' focus on accessible, efficient, and transparent AI methodologies.
