The Agent’s Dilemma: A Comparative Study of LLM-Based Architectures in Strategic Decision-Making

This thesis project revolves around a confrontation between two LLM-based agents faced with a series of simple decision-making scenarios. The first investigated structure is a tool-augmented agent, representing an integration between symbolic and sub-symbolic reasoning. The second is based on an internal dialogue between two simpler agents, simulating cognitive pluralism. Each agent's behaviour is analysed in the context of games of Prisoner's Dilemma under a wider range of conditions, including: payoff configurations, number of rounds, opponent strategy, suggested strategy and underlying model. Results are processed through a combination of metrics from related PD literature and feature analysis. Together, they provide valuable insight into the reasoning process of the proposed architectures.