Agentic AI for QoS-driven Adaptive Routing in Small-to-Medium Scale Software-Defined Networks

Abstract

Software-Defined Networking enables centralized, programmable traffic management, yet existing routing approaches face limitations in balancing multiple quality of service objectives. Classical algorithms minimize single metrics, while Machine Learning and Deep Learning methods require extensive training data and lack explainability. This paper presents an Agentic AI for QoS-Aware adaptive routing in small to-medium-scale SDNs. The proposed approach employs a Large Language Model to reason about traffic requirements and generate optimization weights. The proposed agent, an LLM+LP approach, was evaluated against Classical (Dijkstra), random forest, and LSTM baselines on the NSFNet (14 nodes) and GEANT2 (24 nodes) topologies in a simulated Mininet environment. Across pooled trials from two measurement campaigns (up to 25 per method), the agent achieves quality of service on par with the trained baselines: after Holm correction, no delay, jitter, or packet-loss difference between the agent and any baseline is statistically significant on either topology (Welch ANOVA and Kruskal–Wallis, α=0.05); the agent’s slightly higher mean delay from load-aware path selection does not reach significance. Its distinguishing capability is adaptiveness: in a controlled congestion experiment, the agent is the only method that reroutes traffic in response to sustained congestion, closing the MAPE-K autonomic loop that the static baselines structurally cannot. An ablation against a fixed-weight linear program confirms that this adaptiveness, together with zero-shot deployability and explainable weight generation rather than a raw QoS advantage, constitutes the agent's contribution. The findings indicate that LLM-driven reasoning can substitute for labeled training data in quality-of service routing while providing zero-shot deployability, runtime objective adaptation, and explainable decision-making capabilities that trained models cannot provide.