When cooperation is beneficial to all agents

In a new paper published on arXiv, researchers Alessandro Doldi and Marco Frittelli tackle a fundamental question in multi-agent systems: when does cooperation actually benefit everyone? Their work, 'When cooperation is beneficial to all agents,' establishes a rigorous mathematical framework within general semimartingale theory to analyze exchanges between rational agents. They derive a precise necessary and sufficient condition for the existence of trades—even potentially zero-sum ones—that strictly increase each agent's indirect utility. This condition is characterized by the compatibility between agents' individual preferences and collective pricing measures, bridging concepts from mathematical finance and game theory.

The research provides a formal criterion to determine whether a group of AI agents or economic actors can engage in mutually beneficial cooperation. The framework's strength lies in its generality, applying to both continuous-time models (common in quantitative finance) and discrete-time models (common in algorithmic game theory and multi-agent AI). This work moves beyond anecdotal evidence of cooperation's benefits, offering a testable condition to predict when collaborative strategies will lead to strict Pareto improvements, where every participant is better off without making anyone worse off. It clarifies the precise link between collective market efficiency and the individual rationality of each agent involved.