Agentic TokenCom: A Chain-of-Agents Framework for Multimodal Token Communications

Abstract

In 6G networks, existing semantic communication systems struggle to achieve composable and generalizable end-to-end optimization under dynamic channel conditions and heterogeneous resource constraints. To address this challenge, we propose Agentic Token Communication (TokenCom), a hierarchical Chain-of-Agents autonomous framework composed of a reasoning chain and an execution chain, which transforms TokenCom structure design and instantiation from a static, one-shot configuration into a self-evolving process that is plannable, evaluable, reflective, and iterative. First, we establish an end-to-end multimodal TokenCom system model and summarize three modes of Large Language Model (LLM) token encoding and decoding. We then formulate the operation of Agentic TokenCom as a joint optimization problem of workflow-structure selection and component-instance selection: supported by a memory module, the upper reasoning chain forms a closed loop of planner, evaluator, and reflector agents to generate and continuously refine the workflow structure; under the resulting structural constraints, the lower execution chain dynamically performs model selection, parameter configuration, and interface alignment across different modalities and task requirements. Experimental results demonstrate the effectiveness and feasibility of the proposed framework.