This paper proposes a method for integrating large-scale language models (LLMs) and intelligent services through an agent framework, leveraging the extensive computing resources of 6G networks. LLM-based agents can autonomously plan and act to process diverse environmental meanings and user intents through auxiliary modules and a planning core. However, the limited resources of individual network devices significantly hinder the efficient operation of LLM-based agents, including complex tool invocations. Therefore, efficient multi-level device collaboration is urgently needed. To address this issue, this paper proposes a framework and methodology for an LLM-based multi-agent system with dual-loop terminal-edge collaboration in 6G networks. The outer loop consists of iterative collaboration between a global agent and multiple subagents deployed on edge servers and terminals, enhancing planning capabilities through task decomposition and parallel subtask distribution. The inner loop consists of subagents with dedicated roles that recursively infer, execute, and replan subtasks. Parallel tool invocation generation using offloading strategies is integrated to enhance efficiency. Through case studies on 6G-enabled urban safety management, we verify improved work planning capabilities and work execution efficiency, and thoroughly analyze open challenges and future directions in 6G networks to accelerate the advent of the 6G era.
