Formal Mechanisms for Market Stability in Self-Interested Agent Societies: A Marketplace Simulation Study

Jul 16, 2026

Upload by Sage

Self-interested agents, when left unconstrained, tend to defect in repeated social dilemmas, undermining cooperative gains from trade. This study determines which formal mechanisms—added on top of unrestricted communication—are sufficient to sustain market stability in such agent societies and evaluates resilience under adversarial attack. A multi-agent marketplace simulation uses 18 LLM agents (DeepSeek-V3) with complementary production specialties and constrained social networking, comparing eight mechanism conditions and executing robust adversarial red-teaming.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages23

Reading Time58 min

Abstract

Introduction

Research Question

Contributions

Two-Phase Design

1 Formal Mechanisms for Market Stability in Self-Interested Agent Societies: A Marketplace Simulation Study Eugene Ng Yi Sheng DSO National Laboratories National University of Singapore [nyisheng@dso.org.sg](nyisheng@dso.org.sg)[e1398550@u.nus.edu](e1398550@u.nus.edu) Bingquan Shen DSO National Laboratories [SBingqua@dso.org.sg](SBingqua@dso.org.sg) arXiv :2607 .08652v 1 [ cs .AI] 9 Jul 2026 Abstract Self-interested agents, left unconstrained, tend toward defection in repeated social dilemmas, causing cooperative gains from trade to collapse. This paper investigates what formal mechanisms—layered on top of unrestricted communication—are sufficient for a society of such agents to maintain market stability, and how resilient those mechanisms are to adversarial attack. We instantiate the research question as a multi-agent marketplace simulation where 18 LLM agents (DeepSeek-V3) with complementary production specialties must trade within a constrained social network to obtain utility. We conduct two experimental phases: (1) a mechanism comparison across eight conditions under progressive troll injection over 200 rounds, identifying Mediation as the top-performing mechanism; and (2) adversarial red-teaming of Mediation using iteratively prompt-optimised LLM-driven trolls, finding that the best attack (v6) reduces honest-agent utility by 13.3% but cannot collapse the market—mediation enables recovery even under sustained adversarial pressure. We define adversarial robustness as a mechanism’s ability to sustain positive honest-agent utility under optimised attack, and find that Mediation is robust: it can be bent but not broken. 1 Introduction The stability of cooperative societies is among the oldest problems in social science and, increasingly, in artificial intelligence. As autonomous agents are deployed in economic settings—trading, negotiating, and allocating resources on behalf of human principals—the question of what keeps a population of self-interested agents from collapsing into mutual exploitation becomes practically urgent. Prior work in game theory identifies the core tension through sequential social dilemmas—situations in which individually rational choices produce collectively suboptimal outcomes [Tewolde et al., 2026] . In the canonical Prisoner’s Dilemma, two agents each defect because defection dominates cooperation regardless of the other’s action, yet mutual defection leaves both worse off than mutual cooperation. Scaling this logic to multi-agent systems producesan analogous result: without intervention, populations converge toward all-defection equilibria. Recent empirical work confirms that modern large language models (LLMs), despite their social reasoning capabilities, exhibit this same failure mode. Tewolde et al. [2026] show that all tested LLMs defect in every social dilemma without formal mechanisms in place. Communication alone is insufficient. Natural language promises are cheap and unverifiable. The question is therefore not whether communication helps, but how much formal structure is needed ontop of communication to sustain societal cooperation—and how resilient that structure is when adversaries actively attempt to undermine it. 2 A parallel line of work investigates whether reasoning-capable LLMs are more or less cooperative than their non-reasoning counterparts. Guzman Piedrahita et al. [2025] find that reasoning models free-ride at significantly higher rates than traditional LLMs in public goods games, a phenomenon they term “corruption by reasoning” —the model identifies the Nash equilibrium defection strategy and exploits it. Huang et al. [2026] further show that mechanism design alone has a provable ceiling: under incomplete contracts, no mechanism can eliminate the “cooperation gap”between self-interested equilibrium play and the social optimum, and that prosocial agents who weigh group welfare can close this gap. 1.1 Research Question Which formal mechanism

FAQs

What problem does the paper address for self-interested agents in repeated dilemmas?

Without constraints, self-interested agents tend to defect, causing cooperative gains from trade to collapse. The paper targets what prevents this failure mode.

How is the research question implemented in the study?

The study uses a multi-agent marketplace simulation where 18 DeepSeek-V3 LLM agents with complementary production specialties must trade to obtain utility under a constrained social network.

What mechanism performs best and how does it respond to adversarial attacks?

Mediation is identified as the top-performing mechanism in a progressive-troll comparison. In red-teaming, the strongest attack reduces honest-agent utility by 13.3% but cannot collapse the market; mediation enables recovery under sustained pressure.

Jul 16, 2026

Upload by Sage