Mediation Mechanism Boosts Market Stability in Self-Interested AI Agents

In multi-agent systems where individual agents act in their own self-interest, there's a tendency for cooperation to break down, leading to market instability and reduced overall gains. This research explores various formal mechanisms that, when combined with open communication, can help maintain market stability and resilience in societies of self-interested AI agents. The study used a marketplace simulation involving 18 LLM agents (DeepSeek-V3), each with specialized production capabilities, trading within a constrained social network. The research was conducted in two phases. First, eight different mechanisms were compared under increasing "troll" (adversarial) injection over 200 rounds, identifying "Mediation" as the most effective. Mediation involves an impartial third party facilitating agreements and resolving disputes. The second phase involved red-teaming the Mediation mechanism using iteratively prompt-optimized LLM-driven trolls. Even with the most effective adversarial attack, which reduced honest-agent utility by 13.3%, the Mediation mechanism prevented market collapse and enabled recovery under sustained pressure. The study defines adversarial robustness as a mechanism's ability to maintain positive utility for honest agents under optimized attacks, concluding that Mediation is robust—it can be challenged but not fundamentally broken.