SimCity: Multi-Agent Urban Development Simulation with Rich Interactions

Large Language Models (LLMs) open new possibilities for constructing realistic and interpretable macroeconomic simulations. We present SimCity, a multi-agent framework that leverages LLMs to model an interpretable macroeconomic system with heterogeneous agents and rich interactions. Unlike classical equilibrium models that limit heterogeneity for tractability, or traditional agent-based models (ABMs) that rely on hand-crafted decision rules, SimCity enables flexible, adaptive behavior with transparent natural-language reasoning. Within SimCity, four core agent types (households, firms, a central bank, and a government) deliberate and participate in a frictional labor market, a heterogeneous goods market, and a financial market. Furthermore, a Vision-Language Model (VLM) determines the geographic placement of new firms and renders a mapped virtual city, allowing us to study both macroeconomic regularities and urban expansion dynamics within a unified environment. To evaluate the framework, we compile a checklist of canonical macroeconomic phenomena, including price elasticity of demand, Engel’s Law, Okun’s Law, the Phillips Curve, and the Beveridge Curve, and show that SimCity naturally reproduces these empirical patterns while remaining robust across simulation runs.

💡 Research Summary

The paper introduces SimCity, a novel macro‑economic simulation framework that integrates large language models (LLMs) and a vision‑language model (VLM) to create a multi‑agent environment with rich spatial and market interactions. Four core agent types are modeled: heterogeneous households, firms, a central bank, and a government. Each agent is powered by an LLM that receives a structured prompt containing its personal attributes, recent observations, and a description of the available actions. The LLM generates natural‑language reasoning and returns a JSON‑encoded action, which the simulation engine executes after verification.

SimCity’s environment consists of three interlinked markets that operate on a monthly time step: a frictional labor market, a heterogeneous goods market, and a financial market. Households decide on consumption bundles, labor‑market participation, housing, and financial actions each period. Firms, instantiated from a library of 44 synthetic templates, produce a single differentiated good, hire workers, invest in capital, and set prices using a Cobb‑Douglas production function. The central bank follows a modified Taylor rule to set the policy interest rate, while the government collects bracketed income tax and value‑added tax, then allocates revenue to public‑service construction, universal basic income, or reserves.

A distinctive component is the investment pool coupled with a VLM. Household savings are temporarily pooled; when sufficient capital accumulates, the VLM evaluates macro‑economic conditions and selects a firm template and a geographic location on a rendered city map. This spatial module enables the study of urban expansion dynamics alongside macro‑economic outcomes.

To evaluate the framework, the authors compile a checklist of canonical macro‑economic regularities: price elasticity of demand, Engel’s Law, Okun’s Law, the Phillips Curve, and the Beveridge Curve. Across multiple random seeds, SimCity reproduces these patterns without hand‑crafted rules. For example, Engel’s Law emerges from households’ LLM‑driven consumption decisions that enforce a minimum expenditure on essential goods; the Phillips Curve appears as firms and households adjust wage expectations in response to unemployment rates; and the Beveridge Curve is generated by the matching process in the labor market. The authors also demonstrate robustness to exogenous shocks such as tax hikes and interest‑rate changes, showing responses consistent with standard economic theory.

The contributions are threefold: (1) introducing LLM‑driven agents to macro‑economic simulation, thereby allowing rich heterogeneity and transparent reasoning; (2) providing a visualized, spatially explicit urban environment via a VLM, which links economic activity to city growth; (3) establishing a systematic evaluation methodology based on a comprehensive checklist of macro‑economic phenomena.

Limitations are acknowledged. The behavior of agents heavily depends on prompt engineering, and the paper does not quantify prompt‑induced bias. The monthly granularity omits high‑frequency financial dynamics, and policy agents (government, central bank) still rely on fixed analytical rules rather than fully LLM‑generated policy formation. Future work should explore automated prompt optimization, integration of real‑world GIS and economic data, and quantitative validation against empirical macro‑economic time series.

In sum, SimCity represents a pioneering effort to fuse LLMs and VLMs with agent‑based macro‑economic modeling, opening new avenues for interdisciplinary research across economics, urban planning, and artificial intelligence.