Agentic latency control and cooperative vehicle coordination in 5G-MEC: A prescriptive and explainable AI framework

Abstract

The problem of supporting ultra-low latency in 5G-MEC vehicular networks requires moving beyond passive forecasting toward real-time, proactive traffic management. Current solutions often rely on computationally expensive deep learning predictors or centralized network optimization, yet they frequently stop at prediction and do not complete the vehicle-level prescriptive action loop. This paper proposes an Agentic AI Coordination Framework that links latency-state prediction to immediate, explainable, and cooperative mitigation actions at the vehicle edge. At the core of the framework is a lightweight Support Vector Machine (SVM) trained on the 5G-V2N communication dataset, which enables highly reliable latency_category classification (Accuracy = 0.97; Weighted F1-score = 0.97; High-latency class Precision = 1.00 and Recall = 1.00). The predicted latency state then activates an in-vehicle Agentic AI module that generates dashboard-level prescriptive outputs, including status alerts, action plans, and optimized routing guidance. In addition, the framework leverages 5G-MEC cooperative broadcast to disseminate rerouting alerts to neighboring vehicles, supporting distributed congestion avoidance. To ensure transparency and controllability, the system integrates LIME-based explanations and actionable counterfactual analysis to indicate the minimal feature-level adjustments required to transition between High, Medium, and Low latency states. Overall, the framework forms a unified perception–decision–action pipeline that connects low-overhead inference, rule-based prescriptive control, and cooperative vehicle coordination. To align with sustainable computing and the Special Issue focus on energy-efficient routing in 5G-enabled VANETs, the framework couples low-complexity on-board inference with MEC-assisted broadcast to support real-time coordination while constraining communication and compute overhead. Keywords 5G-MEC, Agentic AI, LIME, SVM, XAI