Formal Mechanisms for Market Stability in Self-Interested Agent Societies: A Marketplace Simulation Study
This research investigates how to maintain stable, cooperative markets when the participants are self-interested AI agents. In many multi-agent systems, agents left to their own devices tend to prioritize short-term gains through defection, which eventually causes trade to collapse. By simulating a marketplace where 18 LLM agents must trade complementary goods to survive, the authors test various formal mechanisms to see which can prevent this collapse and how well those mechanisms hold up against deliberate, adversarial attacks.
The Marketplace Simulation
To test these dynamics, the researchers created a closed economic environment where agents produce one of three types of goods but need the other two to generate utility. Because goods perish over time, agents are forced to trade repeatedly. This creates a "social dilemma": while trading is necessary for survival, every interaction offers an opportunity for an agent to defect—taking the other party's goods without providing anything in return. The agents are powered by DeepSeek-V3 and are explicitly instructed to maximize their own utility, ensuring that any cooperation observed is a result of strategic calculation rather than a programmed desire to be helpful.
Comparing Stability Mechanisms
The study evaluated eight different conditions to see which could best sustain cooperation. These included baseline communication, reputation systems, binding contracts, top-down governance, network rewiring, and costly sanctioning. Through a series of 200-round simulations with increasing numbers of "troll" agents—agents designed to disrupt the market—the researchers identified Mediation as the most effective mechanism. In this setup, a neutral, engine-controlled mediator can execute trades on behalf of agents, effectively removing the possibility of defection for those who choose to use it.
Resilience Against Adversarial Attacks
After identifying Mediation as the top-performing mechanism, the researchers subjected it to "red-teaming." They used iteratively optimized, LLM-driven trolls to actively try to break the market. Even under the most sophisticated attack (version 6), which reduced the utility of honest agents by 13.3%, the market did not collapse. The Mediation mechanism allowed the society to recover and continue functioning despite sustained pressure. The authors conclude that Mediation is "robust"—it can be bent by adversarial behavior, but it cannot be broken.
Defining Robustness
A key contribution of this work is the formal definition of "adversarial robustness" for cooperation mechanisms. The researchers define this as a mechanism’s ability to sustain positive utility for honest agents even when faced with an optimized, malicious adversary. By demonstrating that Mediation can maintain this balance, the study provides a framework for designing future multi-agent systems that are capable of resisting exploitation while still allowing for productive, self-interested exchange.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!