NetIntent: Leveraging Large Language Models for End-to-End Intent-Based SDN Automation

Intent-Based Networking (IBN) often leverages the programmability of Software-Defined Networking (SDN) to simplify network management. However, significant challenges remain in automating the entire pipeline, from user-specified high-level intents to device-specific low-level configurations. Existing solutions often rely on rigid, rule-based translators and fixed APIs, limiting extensibility and adaptability. By contrast, recent advances in large language models (LLMs) offer a promising pathway that leverages natural language understanding and flexible reasoning. However, it is unclear to what extent LLMs can perform IBN tasks. To address this, we introduce \boldsymbol {IBNBench}, a first-of-its-kind benchmarking suite comprising eight datasets: Intent2Flow-ODL, Intent2Flow-ONOS, Intent2Flow-Ryu, Intent2Flow-Floodlight, FlowConflict-ODL, FlowConflict-ONOS, FlowConflict-Ryu, and FlowConflict-Floodlight. These datasets are specifically designed for evaluating LLMs performance in intent translation and conflict detection tasks within the industry-grade and research-focused SDN controllers such as ODL, ONOS, Ryu, and Floodlight. Our results provide the first comprehensive comparison of 33 open-source LLMs on IBNBench and related datasets, revealing a wide range of performance outcomes. However, while these results demonstrate the potential of LLMs for isolated IBN tasks, integrating LLMs into a fully autonomous IBN pipeline remains unexplored. Thus, our second contribution is \boldsymbol {NetIntent}, a unified and adaptable framework that leverages LLMs to automate the full IBN lifecycle, including translation, activation, and assurance within SDN systems. NetIntent orchestrates both LLM and non-LLM agents, supporting dynamic re-prompting and contextual feedback to robustly execute user-defined intents with minimal human intervention. Our implementation of NetIntent across ODL, ONOS, Ryu, and Floodlight achieves a consistent and adaptive end-to-end IBN realization.