Economy of Minds: Emerging Multi-Agent Intelligence with Economic Interactions

Can Markets Solve the Coordination Problem in AI?

Instead of a central boss telling AI agents what to do, this system lets them compete in a market. Agents bid money to perform tasks. Those that do a good job get rich and "reproduce." Bad agents go bankrupt. This allows a group of simple, limited agents to work together to solve much harder problems than any single agent could alone.

In the field of multi-agent systems, engineers struggle with a fundamental tension. How do you get specialized AI models to collaborate without a master controller? Current state-of-the-art approaches usually rely on a central orchestrator (a single coordinating model) to assign roles. However, as the number of agents grows, this central node becomes a performance bottleneck. It also creates a single point of failure. Researchers seek "emergent" intelligence. This is where coordination arises spontaneously from the bottom up.

Beyond the Centralized Orchestrator

The authors of "Economy of Minds" (EOM) investigate how a population of agents can self-orchestrate without centralized control. They want to bypass the need for engineered coordination protocols. The authors ask if simple economic signals—like prices and wealth—can act as a proxy for complex instructions.

Historically, the field has relied on rigid architectural designs. Developers often build a fixed pipeline. Agent A does research, Agent B writes a draft, and Agent C verifies facts. This lacks flexibility. If the task changes, the entire pipeline must be re-engineered. Furthermore, centralized systems struggle with "credit assignment" (the process of determining which agent contributed to a result). Without a clear way to reward the right participants, the system cannot effectively learn which behaviors are useful.

An Economy of Agents

The researchers developed the EOM framework. It treats agent interaction as a marketplace. The investigators moved from "instruction-following" to "incentive-following." In this setup, agents do not receive direct orders. Instead, they participate in auctions to win the right to act .

The mechanism relies on three economic pillars. First, agents compete via auctions. An agent only acts if its "wake-up condition" (a trigger based on the current environment) is met. It must then place a bid to win control. Second, the authors implement a "bucket-brigade" transfer rule .

When a winning agent acts, it pays its bid to the agent that acted previously. This creates a flow of value that travels backward along successful trajectories. Much like a relay race where runners pass a baton, the "baton" here is the currency. This rewards agents for setting up their successors for success. Third, the population undergoes economic selection. Wealthy agents are mutated to create new, similar agents (exploitation). Bankrupt agents are replaced by new, exploratory ones.

The researchers tested this economy across five digital domains. These included mathematical reasoning and accelerator design. They did not start with "smart" agents. They initialized the system with "partial" agents. These models were intentionally restricted in their tools or output length. The goal was to see if the economic structure could compensate for individual weakness.

Emergent Specialization and Success

The study finds that these economic dynamics drive the evolution of collective intelligence. On the MATH benchmark, the authors report that EOM improved Llama-3.1-8B agents from 15.9% to 57.0%. This outperformed the "complete" agent baseline that had full access to the task interface. In accelerator design, EOM reduced the average energy-delay product (a metric for hardware efficiency) to 39.3. This beat the domain-specific DOSA baseline of 80.2 [Table 1].

The authors demonstrate that this success is not merely due to having "more agents." Through ablation studies (tests where components are removed), they show that removing the auction mechanism or wealth-based selection reduces performance. The economy turns trajectory-level success into population-level intelligence. As seen in, wealth concentrates in successful lineages.

Ineffective agents are systematically pruned through bankruptcy.

Surprisingly, the researchers found that the society learns reusable, structural knowledge. In scientific research tasks, "executor" agents evolved complex reasoning motifs in their prompts [Table 4]. These included checking for symmetry or verifying physical constraints. This suggests the market shapes the very way those winners think.

The Path Toward Autonomous Societies

This work suggests a shift in how we approach AI scaling. We may stop engineering perfect "master" models. Instead, we might design "economic environments" where specialized agents thrive.

This offers two consequences for the field. First, it provides a scalable alternative to centralized orchestration. Each agent operates on local information. Its own bid and its own wealth. Therefore, system complexity does not explode as more agents are added. Second, it provides a method for automated discovery. In accelerator design, agents independently rediscovered efficient hardware-software design patterns. They did this without being given any explicit templates.

However, the paper notes a critical limitation. Adaptation currently happens only in "prompt space" (modifying text instructions) with a frozen model backbone. This means agents can learn better strategies. They cannot learn entirely new fundamental representations or cognitive abilities.

If these findings hold, the next step is to move beyond prompt-based evolution. Researchers could test if economic incentives can drive parameter-efficient fine-tuning (updating a model's weights efficiently). If agents can "buy" weight updates, we might witness a truly self-evolving digital civilization.

Figures from the paper