Application of LLM-Based Multi-agent Systems in Business Process Automation

This paper analyzes the integration of large language models and multi-agent systems in business process automation and proposes a theoretical model for cross-agent collaboration. Through literature analysis of core journal papers published between 2019 and 2024, it examines the organizational logic, role division, and collaboration mechanisms of multi-agent systems, identifying deficiencies in dynamic task allocation, state consistency maintenance, and responsibility attribution. It also constructs a tripartite "Cognition–Collaboration–Governance" framework that integrates LLM-driven behavior modeling, multi-agent collaboration, and human-machine co-governance. The system comprises an orchestrator agent, functional sub-agents, and a trusted audit module, employing game-theoretic conflict resolution and dynamic task allocation algorithms to enhance decision transparency and execution efficiency. Tests show that the system performs well in task completion rate and response latency under normal scenarios, but collaboration efficiency decreases in cross-organizational data silo environments. This paper emphasizes the critical role of governance structures and points out the need for interdisciplinary cooperation to promote unified evaluation standards.