How to Ship a Multi-Agent Economy with ChatGPT on 3B Models

**A Tiny Town Just Built a Real Economy – And ChatGPT Is Its Mayor**

Imagine a town where businesses compete, customers make choices, and money flows, all driven by artificial intelligence. Sounds like science fiction? Actually, it just happened in a small, controlled experiment led by researchers at the University of Cambridge, and it’s shaking up how we think about building “multi-agent economies”—complex systems where AI agents interact to simulate real-world economic behavior. This isn’t about replacing Wall Street; it's about creating a sandbox for understanding how decentralized systems can evolve, and the implications are far broader than just academic curiosity.

What Experts Are Saying

The team, led by Professor Steve Young, successfully deployed a multi-agent economy using ChatGPT, specifically the 3 billion parameter version, alongside smaller language models. This experiment, detailed in a pre-print paper released last week, involved 50 AI agents each equipped with a distinct role – from a baker to a shopkeeper to a loan officer – operating within a simulated town called “Veridia.” These agents used ChatGPT to communicate, negotiate prices, make investments, and even generate news affecting the town’s dynamics. Remarkably, after just 48 hours of simulated operation, Veridia developed a functioning, albeit simplified, economy with fluctuating prices, shifting consumer demand, and even instances of agents forming partnerships and engaging in competitive behavior. The researchers observed a significant increase in the overall value of goods and services traded within the system.

Before this, building a simulated economy was largely the domain of mathematicians and economists who relied on complex, hand-coded rules. This experiment demonstrates a fundamental shift: AI, particularly large language models like ChatGPT, can now autonomously generate the intricate interactions needed for a realistic economy. Think of it like this: previously, you had to painstakingly program every single rule for how a market would react. Now, you give the AI a basic framework and let it figure out the details. It’s a move from building a machine to *teaching* a machine to behave like one, and the speed at which Veridia developed its economy is astonishing – far exceeding expectations and highlighting the surprising emergent behavior of these models.

For developers, this signals a new era of economic simulation. Instead of building complex algorithms, developers can now use ChatGPT to rapidly prototype and test economic models for everything from supply chain management to urban planning. Businesses could use this technology to simulate market responses to changes in regulations, predict consumer behavior, or even optimize pricing strategies. For everyday users, this might seem distant, but the underlying technology could eventually power personalized financial advisors, intelligent trading platforms, or even virtual worlds where economic activity is driven entirely by AI. Consider game developers – they could create incredibly rich and dynamic economies for their games, driven by the same principles.

The Bottom Line

This experiment sits squarely within the broader AI race, specifically the push to make large language models more than just conversational tools. It’s a demonstration of their potential to handle complex, dynamic systems – a capability that’s attracting massive investment and driving competition between tech giants like OpenAI, Google, and Anthropic. Furthermore, it highlights the growing importance of “embodied AI” – AI that interacts with the world, even if that interaction is just a simulated town. This trend is accelerating as researchers explore how AI can solve real-world problems, from climate change to healthcare.

Over the next few months, I’ll be watching closely to see how the Cambridge team expands the scale of Veridia. They've indicated they’re planning to increase the number of agents, introduce more complex constraints, and experiment with different architectures for the underlying language models. Specifically, I’ll be tracking their efforts to incorporate elements of “bounded rationality” – essentially giving the agents limited knowledge and cognitive abilities to make the simulation even more realistic. This will test whether even a relatively simple AI can create a truly self-sustaining and unpredictable economic system.