SimCity: Multi-Agent Urban Development Simulation with Rich Interactions

TL;DR

SimCity is a novel multi-agent macroeconomic simulation framework that replaces rigid, rule-based agents with Large Language Models (LLMs) to model a realistic economy consisting of households, firms, a central bank, and a government. By integrating a Vision-Language Model (VLM) for spatial urban planning, the framework successfully bridges macroeconomics and urban expansion, naturally reproducing canonical economic laws and responding realistically to complex external shocks.

What Problem Does This Paper Address?

Macroeconomists traditionally rely on two paradigms, both of which suffer from severe trade-offs. Dynamic Stochastic General Equilibrium (DSGE) models are mathematically elegant but oversimplify human behavior by assuming a single "representative agent" to keep calculations tractable. Conversely, Agent-Based Models (ABMs) allow for diverse, heterogeneous agents but rely on rigid, hand-crafted decision rules that fail to capture human adaptability.

While recent research has explored using LLMs for economic simulations, these frameworks have been overly simplistic—often omitting critical components like firms, diverse goods markets, taxation, or spatial geography. SimCity addresses this gap by creating a rich, unified environment where LLM-driven agents interact across multiple complex markets within a visually rendered, expanding city.

Methods

The framework models an urban economy through four distinct LLM-powered agent types (implemented using gpt-4o-mini for standard reasoning):

• Households: Make monthly decisions regarding consumption bundles, labor market actions (accepting/resigning from jobs), housing choices, and financial allocations (saving, borrowing, or investing).
• Firms: Utilize a Cobb-Douglas production function to transform input goods into specialized outputs, set prices, manage hiring/layoffs, and make capital investments.
• Government & Central Bank: Institutional agents that manage bracketed income/value-added taxes, distribute universal basic income (UBI), and implement monetary policy via a modified Taylor rule with interest rate smoothing.

Expand

A key novelty is the Investment Pool, which utilizes a Vision-Language Model (gpt-4) to act as a city planner. The VLM assesses a rendered map of the city to strategically determine the geographic placement of new firms based on spatial logic (e.g., clustering factories away from residential areas). The simulation runs in two phases: a move-in phase to study urban expansion dynamics, and a development phase with a fixed population to observe steady-state macroeconomic regularities.

Datasets

• Household Profiles: Initialized using real-world demographic data from the U.S. Census Bureau. Initial cash holdings are modeled using a lognormal distribution estimated via maximum likelihood estimation from the 2023 American Community Survey (ACS) microdata.
• Firm Templates: 44 distinct synthetic firm templates are derived from the Organization for Economic Co-operation and Development (OECD) Input-Output Tables (IOTable) to capture realistic supply-chain interdependencies.
• Evaluation Baselines: The framework's emergent properties are evaluated against real-world historical data from FRED (Federal Reserve Economic Data) and existing simulation baselines (traditional ABMs like LEN and CATS, and the LLM-based EconAgent).

Key Insights and Ideas

• Emergence Without Explicit Optimization: Unlike classical models where agents are hardcoded to mathematically maximize utility functions, SimCity agents exhibit realistic economic behaviors purely through LLM role-play and natural language reasoning.
• The Necessity of Sight: An ablation study revealed that giving the spatial planning agent a visual map is critical. Without visual grounding (using text-based coordinates instead), the city layout degrades into unrealistic linear geometries. With the VLM, the city naturally forms a concentric structure matching the classic Alonso-Muth-Mills urban economic model.
• Rich Market Friction: By including qualitative differentiation in goods and a frictional labor market, the framework can capture complex distributional effects, such as technological unemployment and skill-biased income inequality, which traditional models struggle to simulate.

Key Results

The authors compiled a checklist of canonical macroeconomic "stylized facts" to evaluate the simulation. SimCity successfully and robustly reproduced multiple economic regularities across independent runs:

• Okun's Law & Phillips Curve: It naturally captured the short-run inverse relationship between unemployment and inflation, as well as GDP growth.
• Beveridge Curve: It successfully exhibited the negative correlation between the job vacancy rate and the unemployment rate.
• Engel's Law & Price Elasticity: The simulation correctly showed that the proportion of income spent on food decreases as income rises. Furthermore, it reflected realistic demand elasticity—necessities like food were price-inelastic, while luxury/durable goods like motor vehicles were highly elastic.

Expand

• Response to Shocks: When subjected to an exogenous price shock, the economy demonstrated realistic "price stickiness," with prices gradually returning to long-run equilibrium. Additionally, introducing an "automation shock" (replacing physical labor requirements with IT skills midway through the simulation) successfully simulated a widening, persistent income gap between worker types.

Limitations and Future Work

• LLM Aberrations: Because agent decision-making is powered by LLMs, there is a non-trivial probability of highly aberrant decisions over hundreds of iterations (e.g., altering a price by several orders of magnitude). The authors currently mitigate this using simple heuristic-based checks.
• Simplified Financial Systems: The current framework lacks complex financial markets, such as those for bonds, stocks, or their derivatives.
• Lack of Explicit Optimization: Simulated agents do not employ mathematical optimization techniques in their investing or pricing strategies, which may limit the model's ability to accurately simulate highly complex market dynamics. The authors suggest exploring hybrid architectures that combine LLM reasoning with explicit optimization algorithms in future work.

@misc{feng2025simcity,
  author = {Yeqi Feng and Yucheng Lu and Hongyu Su and Yixin Tao and Tianxing He},
  title = {{SimCity: Multi-Agent Urban Development Simulation with Rich Interactions}},
  year = {2025},
  eprint = {2510.01297},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.01297},
}