This paper presents a structured multi-agent deliberation framework in which 100 large language model (LLM) agents, distributed across three frontier models (Claude Opus 4, Gemini 2.5 Pro, GPT-4o), debated whether the Jamaican musician Bob Marley should receive the Order of National Hero. Each agent was assigned a distinct professional identity, argumentation technique, and expert domain drawn from 20 stakeholder categories. Fifty agents argued for the designation; fifty argued against. The simulation ran across three deliberation rounds with structured cross-examination, sidebar caucuses, and open-floor rebuttals. Eleven of the fifty agents who began with an opposing position changed their vote during deliberation, producing a final tally of 61 to 39 in favor. The paper introduces the LLM Agent Persuasion Index (LAPI), a composite metric for benchmarking model susceptibility to adversarial persuasion in multi-agent settings, with applications to AI safety testing, red-teaming, and regulatory evaluation. Findings carry implications for computational social science, AI-assisted policy analysis, and the design of deliberative AI systems for contested public questions.
Adrian Dunkley (Sun,) studied this question.