Adding a Neutral Third Agent to LLM Deliberation: An Exploratory Case Series of Six Dyadic and Triadic Ethical Deliberations

Abstract Multi-agent systems of large language models (LLMs) are now proposed for deliberative tasks in ethically charged domains, yet little is known about how their composition shapes deliberation. This article reports an exploratory case series of six multi-agent deliberations in which LLM agents, each bound to a fixed instructed persona, decided binary ethical dilemmas in three domains (medical intensive care allocation, a corporate layoff, and an academic department closure), once in a dyadic configuration (an institutional decision-maker and an ethical-minority advocate) and once in a triadic configuration that added a critically neutral facilitator. Three frontier models rotated across the roles, sessions were run under a verbatim Relay Protocol, and the complete transcripts and prompts are publicly archived. Analysis is descriptive throughout. Across the six cases (84 agent turns, 37 rounds), the minority agent in every case accepted the operational decision while maintaining a distinguishable ethical position through an identifiable strategy, yielding a six-item descriptive taxonomy of resistance strategies. A coded measure of the authority agent's pressure intensity was lower in the triadic run in all three domains (mean paired difference −0.131), while three of the study's four pre-defined hypotheses were not supported and deliberation length was domain-dependent. Because the personas instructed the very behaviours under study and each configuration was run once, the series supports no causal or emergent-behaviour claim. Its contribution is a transparent, fully archived set of cases, a design vocabulary for facilitator agents, and a precisely stated attenuation hypothesis for confirmatory work.