Agentic AI CPE: Traffic Steering via Small Language Models and Zero-Shot Forecasting

Software-Defined Wide Area Networks (SD-WANs) have become a key enabler of multi-homed enterprise connectivity, allowing traffic to be dynamically steered across heterogeneous access networks. However, current traffic steering mechanisms remain largely reactive and policy-driven, limiting their ability to anticipate performance degradation and to reason effectively over complex, multi-dimensional Quality of Service (QoS) objectives. This paper introduces a novel agentic AI approach for SD-WAN Customer Premises Equipment (CPE) that elevates traffic steering from reactive control to proactive, context-aware decision-making. By combining foundation-modelbased time-series prediction with lightweight semantic reasoning at the network edge, the proposed approach enables predictive traffic steering that operates out of the box in new environments and adapts to application-specific requirements. Unlike prior machine learning-based solutions, our design avoids site-specific training and remains compatible with the compute and latency constraints of commodity CPE hardware. Through an experimental evaluation on a controlled testbed and real-world network traces, we demonstrate the feasibility of deploying predictive and explainable AI-driven traffic steering directly on edge devices. Our results suggest that agentic AI architectures can play a key role in advancing SD-WANs toward truly self-driving network operation.