Building an Economy of AI Minds
Kempner Institute researchers explore whether market-like rules can help groups of AI agents solve hard problems together
The research team developed an artificial marketplace called Economy of Minds, in which AI agents can learn to coordinate through market-like rules: bidding for the chance to act, earning or losing virtual money based on their contributions, and gradually developing specialized roles.
Image: Adobe stock
At a glance
- Kempner Institute researchers developed Economy of Minds, an AI system in which groups of AI agents use market-like rules to coordinate across tasks in mathematics, finance, scientific research, computer hardware design, and coordination of computing systems.
- The new work supports the idea that multi-agent AI systems may become more capable not only through stronger individual models, but also through better rules for coordinating groups of models.
Many artificial intelligence systems work like solitary assistants. A human user gives an AI model a problem, and the model plans, reasons, and produces an answer on its own. But can AI systems be more successful working in teams and delegating responsibilities, as humans tend to do when they undertake complex work?
“Most progress in modern AI has focused on making individual models stronger, through larger models, more training data, or more computation,” explains Zhenting Qi, a Harvard Ph.D. student, who, along with Kempner Institute Investigator Yilun Du and collaborators has created and tested a new market-inspired system of collaboration between AI agents. “[We] wanted to explore a complementary way to scale intelligence: can a population of individually limited agents organize into a system that is substantially more capable than any single member?”
The answer, it seems, based on the new study, is yes. The research team developed an artificial marketplace called Economy of Minds, in which AI models, or “agents,” can learn to coordinate through market-like rules: bidding for the chance to act, earning or losing virtual money based on their contributions, and gradually developing specialized roles.
In tests across math, finance, scientific research, computer chip design, and distributed-system optimization, these agent societies improved with experience. In several tasks, groups of agents with limited abilities outperformed a more powerful AI system working alone.
Importantly, rather than having a human engineer or central controller assign every role, the system uses bidding and rewards to help agents learn when their contributions are useful.
“One hope is that a lot of the intelligence we see in human society is actually social,” said Du, an assistant professor of computer science at the Harvard John A. Paulson School of Engineering and Applied Sciences (SEAS) and the senior author of the study. “Whenever you look at the evolution of a field, it is never one person driving the field forward. Instead, many groups of people form, and that drives broad, macro-level ideas.”
Using competition to enable collaboration
In the case of Economy of Minds, performance improvement did not come simply from diving up labor among multiple agents. Rather, when the researchers removed elements of the market design, such as the auction system, performance generally fell, suggesting that the system’s organization, rather than the number of agents alone, contributed to the performance gains.
Economy of Minds uses incentives, rather than a predetermined workflow, to decide which agents contribute at each stage of a task. Instead of assigning each agent a fixed role, the researchers give the agents a set of market-like rules and allow the system to determine which agents should contribute.
“We do not need humans to predefine the collaboration structure,” explains Chloe Huangyuan Su, a Kempner Graduate Fellow and the second author of the study. “We also do not need a centralized AI controller to manage the agents.”
The system coordinates agents through a relatively small set of simple market-like rules.
“What excites me most is the gap between the simplicity of the mechanism and the complexity of the behavior it produces,” Qi said. “The agents follow a relatively small set of economic rules, yet coordination, specialization, and division of labor emerge from their interactions.”
A market for AI agents
The market-like structure of Economy of Minds operates through auctions. At each stage of a task, agents estimate whether they are well suited to contribute and place bids accordingly.
These auctions rely on a form of virtual money that allows the system to track which agents are useful in contributing to a given task. Agents whose actions help complete a task earn money and gain more opportunities to act. Agents that repeatedly take irrelevant or unhelpful actions lose money and may eventually be removed from the system.
The goal of the system is not to fully replicate a real-world economy. Instead, the researchers treat prices as signals that indicate which agents are best suited to provide help.
“One way to think about an economy is that you have many agents with partial information,” Du said. “Prices allow for the exchange of information about which agents have useful tools, and they teach agents to specialize and act only when they have the strongest competitive advantage.”
Future economies of AI agents
So far, the team has tested relatively small societies of 10 to 17 AI agents. The agents used existing language models, with their roles largely defined through prompts.
On a set of math problems, the accuracy of a group of specialized agents built with the Llama 3.1 language model increased from 15.9 percent initially to 57 percent after the system adapted through a series of practice problems. The adapted multi-agent system’s 57 percent accuracy exceeded the 51.9 percent achieved by a single agent built with the same language model. The system also performed well on tasks involving financial and scientific research, computer hardware design, and the coordination of computing systems.
The researchers are now considering larger versions that could support more sophisticated forms of collaboration and specialization. Du said future societies might include agents that improve with experience, agents that process images, and agents that interact with the physical world.
Qi says that he and his colleagues ultimately want to determine the conditions under which groups of agents provide greater benefits than a single stronger model. They plan to compare three approaches with equivalent computing budgets: training a stronger model, giving a single model more computation for each problem, and dividing the work among a larger group of agents.
Su is interested in another step: moving from “economies of minds” to “institutions of minds.” In the current system, humans still design the rules of the AI economy. In a future version, agents might evaluate their own performance, identify which rules help or hinder collaboration, and propose revised rules for future tasks or versions of the system.
The study suggests that AI systems may become more capable through better organization, rather than simply through more powerful individual models. By allowing groups of agents to learn when to compete, when to cooperate, and how to divide complex work, Economy of Minds offers a framework for studying how collective intelligence may emerge in AI systems.
