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

The paper presents a low-cost alternative to running thousands of expensive LLM agents: query a teacher model such as DeepSeek a few hundred to few thousand times, fit each agent with a two-to-twelve-parameter behavioral surrogate, and simulate the resulting society on a laptop. Its central contribution is an interaction-order-by-memory taxonomy that predicts when coarse-graining works. Global feeds produce mean-field behavior, with surrogate error decreasing as N^-1/2; community feeds require block or graphon closures and leave an O(1) floor; local, long-memory systems require pair approximations and memory kernels. Across eight named simulations, including EconAgent, OASIS, AgentSociety, LLMTraveler, Generative Agents, and epidemic models, the predicted error trends held. In EconAgent, the surrogate reproduced the behavioral Phillips correlation at -0.569, while showing that the reported Okun relationship is largely an accounting identity. A 2×2 ablation found that reasoning, rather than prompt wording, shifts the Phillips effect by 0.50 correlation points. On a genuine DeepSeek response function, curvature created a private-feed error floor of 0.018, demonstrating when Jensen bias defeats averaging. The method also transfers beyond LLM societies: a degree-aware epidemic closure achieved RMSE 22 in 0.34 seconds, versus RMSE 48 and roughly 400 seconds for a differentiable GPU model. Cross-model tests spanning DeepSeek, OpenAI’s GPT-4o, Anthropic, Google, Meta, and Llama suggest that perception structure is broadly shared, while response curvature and reasoning remain model-specific.