Toward Autonomous Network Management: An Intelligent and Secure Agentic Framework for Network Configuration

The increasing scale and complexity of network infrastructures demand efficient and secure automated configuration solutions. Existing approaches often face challenges such as incompatibility with devices from different manufacturers, limited flexibility in network resource allocation, high demands for personnel expertise, and insufficient security in configuration content. To address these issues, we propose an innovative agent-based framework that leverages large language models (LLMs) to achieve secure and intelligent network configuration. The framework integrates a multi-agent system, which includes: an intent recognition and rewriting agent that uses the Chain of Thought (CoT) method to accurately interpret user intent; a self-reflection retrieval agent that combines hybrid retrieval methods with self-reflection mechanisms to optimize knowledge retrieval; a security agent specialized in network security verification; and a content integration agent that generates final configuration files. Experimental results on a custom dataset demonstrate transformative improvements in accuracy, efficiency, and security. Our framework substantially outperforms both original large language models and strong Retrieval-Augmented Generation (RAG) baselines, with metric scores such as CodeBLEU and BLEU increasing by over 44% compared to the RAG baseline.