Rehearsing Urban Futures: Generative Agents and Daily Life Narratives

Urban planning has long been a conversation between two different languages. On one side, experts speak in technical maps, zoning codes, and mathematical models that predict how a city might grow. These tools are powerful for calculation, but they are often difficult for ordinary people to read or understand. On the other side, residents speak in stories about their daily lives: the bus they wait for, the shortcut they take to the grocery store, or the fear they feel walking down a dark street. When these two groups meet in a community meeting, the technical tools often dominate the discussion, forcing residents to react to immediate, small problems like a broken streetlight rather than imagining how the whole neighborhood could change over the next generation. This gap leaves many big questions about housing, climate, and fairness outside the room. Researchers at the Massachusetts Institute of Technology have developed a new way to bridge this divide, using computer programs that can listen to a person's story and then act it out in a virtual city, allowing everyone to see and debate the future before it is built.

The core idea behind this work is to turn a city simulation into a place for rehearsal rather than just prediction. Instead of trying to guess exactly what will happen, the system creates a space where residents can test their ideas and see the consequences. The researchers built two different versions of this system to see how people interact with them. The first version, called the Neighborhood Platform, was set up in San Francisco's Inner Richmond district. Here, the team took real interviews with local residents and turned them into three distinct characters: a teacher, a small business owner, and a long-time neighbor. Each character had a specific daily routine based on what they told the researchers. The system then created a digital map of the neighborhood and populated it with these characters, showing how they moved through the streets. The installation used large screens to show two views side by side: one showing the neighborhood as it is today, and the other showing what it would look like if the residents' wishes for better transit or more shops were granted. This allowed groups of people to watch how a change in one part of the street might affect the daily life of a neighbor they had never met, fostering a deeper understanding of how policies impact different people.

The second version, the Resident Platform, flipped the script. Instead of watching a character based on someone else's story, the user became the character. A resident sat down with a computer, answered questions about their own commute and daily habits, and the system built a digital version of them, known as an agent. This agent was not just a static picture; it was a program that could think and move. The user could then pause the agent's walk at any point on their route and suggest a change in plain English, such as "remove the cars" or "add a cafe." The system would instantly edit a photo of that street corner to show the change and then watch how the agent reacted. If the new street looked safer or more pleasant, the agent might choose to walk there instead of taking the usual detour. The user could see the agent's reasoning written out, explaining why it made a certain choice, and could agree or argue with that logic. This process turned the planning session into a dialogue where the resident could correct the computer if it misunderstood their needs.

To test how well this worked, the researchers brought the Resident Platform to workshops in San Francisco and Boston, working with thirty residents who knew their neighborhoods well. They found that the most valuable moments happened when the computer got something wrong. When the digital agent made a choice that the real person knew was impossible or unsafe, the resident would stop and explain why. For example, if the agent walked down a street that was poorly lit, the resident would point out that they would never go there because of a past incident. This correction provided the researchers with deep, nuanced knowledge that a simple survey could never capture. The study showed that the computer did not need to be perfectly accurate to be useful; it only needed to be close enough to provoke a reaction. The agent acted as a bridge, or a shared object, between the resident's memory and the planner's model. It was a tool that could be argued with, forcing the system to learn from the human's specific experience.

The results suggest that this approach can shift public participation from a passive complaint session into an active design process. When residents saw their ideas visualized and tested against the daily routines of others, their conversations moved beyond small fixes like adding a bench. They began to discuss complex, long-term futures, such as what happens to a neighborhood if cars are removed entirely. The researchers noted that while the digital agents were good at understanding what people cared about, they were less perfect at predicting exactly how people would behave, often relying on general assumptions rather than personal quirks. However, this imperfection was part of the process. The system was designed to be a draft, not a final verdict, inviting residents to edit their own reality. By making the simulation a place for debate rather than a black box of expert knowledge, the tools helped residents feel more involved and capable of shaping the complex systems that govern their cities. The work concludes that the future of urban planning may not lie in more perfect predictions, but in better ways for communities to rehearse and refine the futures they want to live in.