TaskON: Task-Oriented Networking for Agentic AI

In sensor-cloud networks, deploying large-scale models such as Large Language Models (LLMs) for real-time decision-making is often limited by latency and resource constraints. A promising solution is to enable collaborative operation between large models that excel at complex reasoning and lightweight models optimized for specialized, resource-efficient tasks. In this paper, we present a novel Agentic AI framework where specialized agents assume distinct functional roles within a coordinated system. We validate its effectiveness in intelligent decision-making and control through emergency traffic evacuation scenarios. Specifically, we first introduce the integration of the Model Context Protocol (MCP) to incorporate lightweight expert modules via a standardized, context-aware interface, enabling agents to make more informed decisions in rapidly evolving environments. Furthermore, to address the high-frequency communication demands inherent in real-world agentic systems, we propose a Task-Oriented Networking (TaskON) architecture that decouples information flows between agents and minimizes communication latency. Experimental results demonstrate that our proposed framework significantly enhances evacuation efficiency and system responsiveness compared to traditional approaches, achieving up to a 35.5% faster evacuation and about a 70% lower average communication latency.