Requirements engineering for healthcare systems is inherently complex, primarily due to the divergence in objectives between clinical stakeholders and technical architects, compounded by the limited availability of medical experts for extended consultations. Traditional approaches to requirements elicitation and negotiation face significant limitations in safety-critical domains. This paper presents MedNegotiator, a novel multi-agent framework that overcomes these limitations by integrating generative AI and large language models with formal value quantification and game-theoretic resolution mechanisms. To the best of our knowledge, MedNegotiator is one of the first to combine persona-driven LLM agents, QFD-derived utility functions, Nash bargaining with safety-critical guardrails, and a formal Consistency Guard within an automated negotiation process tailored to healthcare requirements engineering. The framework employs multiple agents representing clinical and technical perspectives that negotiate over utility functions. Conflicts are detected and resolved iteratively, producing Pareto-optimal requirements that balance clinical value and technical feasibility. We formalize the negotiation objective and provide analytical and empirical evidence that, under fixed utilities and hard constraints within a negotiation episode, the architecture converges to a safety-feasible Pareto-efficient agreement. Ambiguities are reduced and requirements are prioritized according to weighted scores. Evaluation demonstrates effective conflict resolution, high convergence rates, and direct applicability. Results achieve 95% agreement within 120 negotiation steps, with 80% of outputs adopted without modification. Human oversight ensures safety and mitigates residual risks. A complexity and runtime profile confirms operational feasibility at realistic scales, with end-to-end cost dominated by LLM inference while symbolic and numeric components remain lightweight.
