An Agentic Approach For Dynamic Software-Defined Network Management Using Large Language Models

In recent years, Large Language Models (LLMs) have been in the spotlight over very diverse fields of research and development. Powered by the concept of Agents, LLMs have acquired substantial tools to perceive and act upon the environment they are applied to. Even though a few works have explored the application of LLMs in the field of dynamic management of network environments, most stop at the application of this technology at Natural Language Processing tasks such as identifying intents, leaving the interpretation and execution of those to humans. This work proposes the application of a reflexive architecture for LLM-powered Agents, enabling perception of network and application data and aiming to autonomously act upon a Software-Defined Network (SDN) environment. The problem of dynamic SDN management and application-aware optimizations is tackled by the novel approach in this work. An experiment was conducted with the proposed Reflexive Agent acting upon a simulated 5G Core Network and video streaming application, through which early results show that the applied LLM-powered Agent approach was able to exert precise actions, achieving the defined intents and providing comprehensive reasoning to support its actions.