When AI Agents Build a Town, Money Stays Put

City SimulationsAI & Machine LearningAI Urban Planning

Imagine a town run entirely by AI agents that never spend the money they earn—this study uncovers why.

TL;DR
• LLM agents in a simulated town economy fail to transmit money to wages, even after tourist shocks.
• Wealth distribution appears frozen only because studies stop too early; longer runs show gradual change.
• The choice of LLM model, not memory, drives these outcomes, highlighting the need for better economic incentives in AI systems.

Summary
A closed, money‑conserving simulation placed one hundred memory‑equipped LLM agents in Pokhara Lakeside, Nepal, to run businesses, earn wages, and set prices over up to 26 simulated weeks. The platform logged every transaction, tool call, and price change, enabling exact ledger reconciliation and independent double‑validation of all 2.44 million decisions and 21.5 billion tokens.
• Run long‑horizon simulations (12 + weeks) to detect wealth‑rank drift that short runs miss.
• Prioritize model selection and prompt design, as the backing LLM drives macro outcomes more than agent memory.
• Incorporate explicit wage‑adjustment or price‑update mechanisms to break the observed transmission failure.
• Release full transaction logs and validation pipelines to support reproducibility and external audit.
• Monitor tool‑failure rates, especially for social coordination actions, and adjust system caps to avoid hidden bottlenecks.

Tourist demand shocks increased total business revenue by 4.62 ×, decomposed into a 1.50 × rise in the number of active firms and a 3.07 × rise in revenue per firm, yet wages rose only 1.03 × and menu prices changed for less than 0.3 % of items. A randomized cash transfer of NPR 5,000 to twenty agents showed that 96.7 % of the grant remained unspent after 311 pulses, yielding a marginal propensity to consume of 3–4 %, far below human benchmarks.

Wealth‑rank persistence fell from a Spearman ρ of 0.964 after two weeks to 0.752 after twenty‑six weeks, with the share of agents changing wealth quintile rising from 21 % to 46 %. Model swaps (Qwen 3.8‑27B to gpt‑oss‑20b) altered all macro metrics significantly, while removing memory had no detectable effect. Social tools such as invitetotalk failed over 94 % of calls, accounting for 80 % of total tool failures, whereas economic tools succeeded above 95 %.

The study highlights a systemic failure of monetary transmission in LLM‑driven economies, emphasizing the need for transparent reporting, rigorous validation, and stakeholder‑informed design of incentive mechanisms before real‑world deployment.

If we want AI‑run economies to mirror human ones, we must first teach agents to move money, not just hold it.