MAFIA-NeT: Multi-Agent Framework for Interactive Agricultural Negotiation and Trading Systems

Dung T. Phan, Chi N. L. Phan, Long S. T. Nguyen, Phuc T. Dao, Quan M. Bui, Tin T. Ngo, Thi T. Nguyen, Tho T. Quan

Abstract: Large Language Models (LLMs) have demonstrated strong reasoning capabilities, yet they remain limited in sustained coordination and reliable real-world operation. These limitations have motivated the development of LLM-powered agents and LLM-based Multi-Agent Systems (MAS), which are better suited for domains that require adaptive reasoning and iterative decision-making. E-Agri commerce is one such domain, as agricultural negotiation involves dynamic pricing, diverse buyer requirements, and rapidly changing market conditions. Existing approaches mainly focus on prediction rather than interactive negotiation, making them insufficient for multi-turn bargaining scenarios. To address this gap, we propose MAFIA-NeT, a multi-agent framework for end-to-end agricultural trade negotiation. The system coordinates specialized agents for structured data parsing, market-informed price reasoning, and strategic decision support. A key contribution is the LLM-Guided Negotiation Subspace (LGNS), whic