Sarcouncil Journal of Engineering and Computer Sciences

Abstract: The financial service sector faces significant challenges in orchestrating multiple specialized software agents to handle diverse customer interactions efficiently. Traditional static rule-based coordination systems demonstrate considerable limitations when encountering variations in customer inquiry patterns, leading to operational rigidity and increased processing times. This manuscript introduces an innovative self-improving multi-agent architecture that employs reinforcement learning to dynamically optimize inter-agent coordination in financial service environments. The architecture features specialized agents with distinct functional roles, individual policy models, distributed learning capabilities, and an adaptive coordination network. A distinctive contribution lies in the novel reward mechanism that balances individual agent performance with system-wide efficiency through multiple weighted components and temporal credit assignment. Validation through both synthetic and real-world financial service scenarios confirms substantial improvements in resolution time, system throughput, agent utilization, and customer satisfaction scores. The system demonstrates continuous performance enhancement even after initial convergence, confirming the superiority of adaptive learning-based coordination over static orchestration methods for complex financial service operations.