An Expert Multi-Agent System for End-to-End Requirements Engineering: A Neuro-Symbolic Approach with Game-Theoretic Negotiation and Graph-Based Traceability

Requirements engineering (RE) is a complex, knowledge-intensive process that remains a major bottleneck in modern software development due to fragmented elicitation, ambiguous stakeholder intent, and weak traceability. While Large Language Models (LLMs) have shown promise in automating isolated RE tasks, single-agent pipelines struggle to preserve context, negotiate conflicts, and maintain validation-aware traceability. This paper presents an expert multi-agent system designed for end-to-end RE, integrating four specialized agents for elicitation, analysis, conflict resolution, and validation. The system employs a versioned Knowledge Graph (Graph RAG) to supply the disagreement point and search space for a Nash Bargaining negotiation engine, whose outputs are gated by a neuro-symbolic validation loop before being committed back to the graph. Experiments on public benchmarks and proprietary industry datasets demonstrate statistically significant improvements over the strongest baselines across all evaluated metrics. Comprehensive ablation studies confirm the contribution of each architectural component, particularly the versioned knowledge graph, game-theoretic negotiation, and neuro-symbolic validation. The results indicate that the proposed system effectively reduces expert workload and provides a viable foundation for deployment in complex, regulated industrial environments.