This paper explores the integration of large language model (LLM) agents with 6G networks, focusing on perception, grounding, and alignment. The authors propose a split learning system that leverages collaboration between mobile devices and edge servers to deploy LLM agents efficiently. This system distributes multiple LLMs with different roles across mobile devices and edge servers to perform user-agent interactive tasks collaboratively. The LLM agents are split into perception, grounding, and alignment modules, enabling inter-module communications to meet extended user requirements on 6G network functions, including integrated sensing and communication, digital twins, and task-oriented communications. A novel model caching algorithm is introduced to improve model utilization in context, reducing network costs of the collaborative mobile and edge LLM agents. The paper discusses the challenges of deploying LLM agents on mobile devices in 6G networks, including limited device capacity and the need for offloading complex tasks to global LLMs on edge servers. It also highlights the advantages of partitioning LLM agents into mobile and edge agents, such as flexible deployment, long-horizon collaboration, and enhanced adaptability in dynamic environments. The authors present a case study on the application of mobile and edge LLM agents in generating accident reports collaboratively, demonstrating the system's effectiveness in real-world scenarios. The paper also introduces a metric called age of thought (AoT) to assess the significance of thoughts generated by LLMs during reasoning and planning processes. Based on this metric, the Least Age-of-Thought (LAoT) model caching algorithm is proposed to evict global models with the least impactful thoughts, reducing grounding cost in terms of latency, resource consumption, and performance loss for serving edge LLM agents in 6G networks. The authors conclude that the proposed split learning system provides a democratic, flexible, and long-horizon approach for running sustainable AI agents in open-ended environments. They also highlight the importance of further research in integrating 6G networks with AI agents, incorporating next-generation technologies such as multiple access, metasurface, and over-the-air computation to support LLM agents in dynamic wireless environments. Additionally, they emphasize the need to address model privacy concerns during collaboration between mobile and edge LLM agents to prevent potential information breaches.This paper explores the integration of large language model (LLM) agents with 6G networks, focusing on perception, grounding, and alignment. The authors propose a split learning system that leverages collaboration between mobile devices and edge servers to deploy LLM agents efficiently. This system distributes multiple LLMs with different roles across mobile devices and edge servers to perform user-agent interactive tasks collaboratively. The LLM agents are split into perception, grounding, and alignment modules, enabling inter-module communications to meet extended user requirements on 6G network functions, including integrated sensing and communication, digital twins, and task-oriented communications. A novel model caching algorithm is introduced to improve model utilization in context, reducing network costs of the collaborative mobile and edge LLM agents. The paper discusses the challenges of deploying LLM agents on mobile devices in 6G networks, including limited device capacity and the need for offloading complex tasks to global LLMs on edge servers. It also highlights the advantages of partitioning LLM agents into mobile and edge agents, such as flexible deployment, long-horizon collaboration, and enhanced adaptability in dynamic environments. The authors present a case study on the application of mobile and edge LLM agents in generating accident reports collaboratively, demonstrating the system's effectiveness in real-world scenarios. The paper also introduces a metric called age of thought (AoT) to assess the significance of thoughts generated by LLMs during reasoning and planning processes. Based on this metric, the Least Age-of-Thought (LAoT) model caching algorithm is proposed to evict global models with the least impactful thoughts, reducing grounding cost in terms of latency, resource consumption, and performance loss for serving edge LLM agents in 6G networks. The authors conclude that the proposed split learning system provides a democratic, flexible, and long-horizon approach for running sustainable AI agents in open-ended environments. They also highlight the importance of further research in integrating 6G networks with AI agents, incorporating next-generation technologies such as multiple access, metasurface, and over-the-air computation to support LLM agents in dynamic wireless environments. Additionally, they emphasize the need to address model privacy concerns during collaboration between mobile and edge LLM agents to prevent potential information breaches.