Kantian Equilibrium in the Age of Multi-Agent Systems

As autonomous agents increasingly operate in multi-agent systems, game-theoretic questions become practically relevant rather than merely conceptual. Agents interact, compete for shared resources, and pursue locally specified goals; their interaction can therefore be studied as a strategic game. Prior work on LLM agents and game theory has largely emphasized Nash-like behavior, where agents maximize individual payoffs or approximately implement best responses. Similarly, much work in reinforcement learning is centered on optimizing a specific, often greedy, objective. However, Nash equilibrium is not the only economically meaningful solution concept, and it does not fully describe cooperative or norm-sensitive behavior. This limitation is especially important in common-resource environments, where individually rational actions can lead to collectively inefficient outcomes. In this context, tragedy-of-the-commons scenarios remain relatively underexplored for LLM agents, but may become important as multi-agent systems increasingly share computational resources, budgets, tools, memory, and API capacity. We raise the question of whether LLM agents can exhibit behavior closer to Kantian equilibrium, a cooperative solution concept inspired by the categorical imperative: agents evaluate actions by considering what would happen if relevant others acted according to the same rule. We instantiate this question in a tragedy-of-the-commons game motivated by local multi-agent deployments with limited shared computational resources. Following the mixed Kantian--Nashian perspective of prior work, we compare payoff-maximizing agents with agents prompted using Kantian formulations, and test whether introducing Kantian agents improves strategies and payoffs. Our results suggest that Kantian reasoning, although with several qualifications, can be reproduced by LLM agents and may be useful for multi-agent deployment. At the same time, the experiment is sensitive to how the Kantian agent is specified, and produces nontrivial cases in which weaker models appear more effective.