Generative AI Agents Move Into a Simulated City to Tackle Urban Complexity

Cities are the most complicated artifacts humanity has ever built. Beneath the skyline of any metropolis lies a dense web of physical infrastructure, social relationships, economic exchanges, and environmental processes, all interacting in ways that constantly surprise the engineers and policymakers who try to manage them. Traffic congestion, air pollution, housing inequality, and the uneven distribution of public services are not isolated problems but symptoms of this deeper complexity. A new review published in PLOS Complex Systems by Fengli Xu, Jun Zhang, Chen Gao, Peijie Liu, Jie Feng, and Yong Li argues that the field of urban simulation is now at a turning point, and that large language model agents may finally provide the missing ingredient: a way to represent the subtle, adaptive, and deeply human behavior that has always eluded conventional models.

For decades, researchers have tried to capture the pulse of cities with data. The rise of big data analytics, urban computing, and digital twin technology has produced increasingly detailed digital replicas of urban systems, fed by sensors, mobile phone records, transit smart cards, and satellite imagery. These tools can map traffic flows in near real time, predict energy demand, and visualize how a new metro line might reshape a neighborhood. Yet the authors identify a persistent gap between these technological capabilities and their practical impact on real urban problems. The reason, they contend, is that most existing simulations treat people as statistical aggregates or as rule-following particles, missing the complex and often irrational ways in which actual residents decide where to go, when to travel, and how to respond to a changing city.

Traditional agent-based modeling has long recognized this weakness. In an agent-based model, thousands of individual software agents interact on a virtual landscape, and collective patterns such as traffic jams or epidemic spread emerge from the bottom up. The approach has been influential in complexity science, but its agents are typically governed by hand-written rules that are crude approximations of human cognition. A simulated commuter might follow a fixed probability of choosing one route over another, whereas a real commuter weighs habit, weather, a text message from a friend, the price of parking, and a hundred other factors. Building believable agents by hand has proven so laborious that many models sacrifice behavioral realism for computational tractability, limiting their usefulness for policy analysis.

The recent advance of large language models has changed this calculus dramatically. LLMs trained on vast corpora of human-generated text exhibit emergent abilities to simulate human-like behavior: they can plan, reason about social norms, express preferences, and adapt their actions to context in ways that resemble real people far more closely than scripted rules ever could. When researchers give an LLM a persona and place it in a simulated environment, the model generates plausible daily routines, social interactions, and even economic decisions. This capacity, the review argues, presents an important opportunity for characterizing human behavior in urban studies, because the agent no longer needs every quirk to be programmed explicitly; much of the richness emerges from the language model’s internalized understanding of how people live.

To harness this potential, the authors conceptualize a novel research infrastructure they call the Urban Generative Intelligence, or UGI, platform. The core idea is to ground LLM agents in a simulated urban environment rather than leaving them as free-floating conversational systems. In the UGI design, a city simulator emulates the urban environment as a textual space: streets, buildings, shops, transit stops, and other agents are described in natural language, and the state of the world is updated as agents act. Each LLM agent perceives this textual city, reasons about its goals, and takes actions through a natural language interface, which the simulator then translates back into changes in the urban state. The loop closes, and a population of generative agents lives out simulated days inside a working model of a city.

The choice of a textual interface is more than a convenience. Natural language serves as a universal medium through which heterogeneous information about the city, from the layout of a street to the crowd level of a restaurant, can be expressed in a form that language models natively understand. It also lowers the barrier for researchers from different disciplines: an economist, a transportation engineer, and a sociologist can all specify scenarios, inject interventions, and query the simulation in plain language rather than writing bespoke code. The authors position UGI as an open platform for diverse intelligent and embodied urban tasks, meaning that agents can be given bodies and goals, sent to commute, shop, socialize, or evacuate, and their collective behavior observed and measured.

What makes such a platform scientifically valuable is the range of questions it could address. Policy experiments that are costly or unethical to run in the real world, such as closing a major road, changing transit fares, or relocating a hospital, could be tested on a population of generative agents whose behavior adapts in human-like ways. Epidemic models could capture not just contact rates but the social reasoning behind compliance with health measures. Studies of segregation and inequality could examine how individual decisions about where to live and work aggregate into city-scale patterns. Because the agents communicate and decide in language, their reasoning traces can be inspected, offering researchers a window into the micro-foundations of emergent urban phenomena that purely mathematical models cannot provide.

The review situates UGI within a longer lineage of interdisciplinary work on complex urban systems. Urban computing has supplied the data pipelines and computational methods for sensing and predicting city dynamics. Digital twins have supplied the ambition of maintaining living, continuously updated replicas of physical infrastructure. Agent-based modeling has supplied the theoretical framework of emergence, in which macroscopic order arises from microscopic interactions. LLM agents, in this framing, are the cognitive layer that was missing, the component that lets each simulated individual behave less like a particle and more like a person. The authors survey the recent literature across these fields and argue that their convergence is what makes a generative urban simulation platform feasible now, when none of the individual technologies alone could have achieved it.

Significant challenges remain on the road from concept to working infrastructure. Running thousands of LLM agents at city scale demands enormous computational resources, and the cost of querying large language models repeatedly for every decision of every agent is far from trivial. Fidelity is another concern: a language model’s plausible-sounding behavior is not guaranteed to be statistically accurate behavior, and biases embedded in training data could distort simulated populations in ways that are hard to detect. Validating that generative agents reproduce real mobility patterns, consumption choices, and social dynamics will require careful calibration against empirical urban data. There are also questions of privacy and governance, since models trained on human text could inadvertently encode identifiable behavioral patterns of real individuals.

Even with these caveats, the vision articulated by Xu and colleagues marks a striking shift in how scientists may come to study cities. Instead of building models from equations and rules alone, researchers could grow virtual societies of generative agents inside simulated urban environments, observe what emerges, and intervene experimentally before anything is tried on real streets. The Urban Generative Intelligence platform, as described in the review, is a foundational step toward that future: an open testbed where language-model agents, a textual city simulator, and natural language interaction come together to make urban complexity something that can be explored, understood, and perhaps ultimately managed. If the approach matures, the city of the future may first be lived in, one simulated day at a time, by thousands of artificial residents whose collective behavior teaches us how to design better places for real ones.

Subject of Research: A review and platform concept for grounding large language model agents in simulated urban environments to model complex city systems

Article Title: Towards a foundational platform for generative agents in simulated city environment

Article References: Xu, F., Zhang, J., Gao, C., Liu, P., Feng, J., & Li, Y. (2026). Towards a foundational platform for generative agents in simulated city environment. PLOS Complex Systems, 3(3), e0000093. https://doi.org/10.1371/journal.pcsy.0000093

Image Credits: AI Generated

DOI: 10.1371/journal.pcsy.0000093

Keywords: generative agents, large language models, urban simulation, digital twins, agent-based modeling, urban computing, complex systems, smart cities, Urban Generative Intelligence, human behavior simulation, city digital twin, LLM agents