Multi-Agent Debate System for AI-Based Decision-Making: A Framework for Enhanced Reasoning Through Collaborative Intelligence

Single-agent Large Language Models (LLMs) demonstrate limitations in complex decisionmaking scenarios, including domain-specific bias, overconfidence, and inability to integrate diverse perspectives. This paper presents the MultiAgent Debate System (MADS), a collaborative AI architecture leveraging specialized agents to generate robust insights through structured argumentation. Implemented using CrewAI framework with Llama 3 models via Groq's LPU infrastructure, MADS orchestrates three specialized agents (Advocate, Critic, Judge) in sequential debate workflows. Testing on interdisciplinary datasets demonstrates 73% improvement in argument quality over singleagent baselines, with average response generation under 8 seconds. The system produces multi-format outputs (transcripts, summaries, PDF reports) accessible to nontechnical users. By replicating human deliberative processes through agent-based debate, MADS advances interpretable, transparent AI decision-support systems while addressing critical gaps in crossdomain reasoning and perspective integration.