Structured Agentic Intelligence: Enabling Explainable and Resilient UAV Operations in Low-Altitude Wireless Networks

Abstract

Future low-altitude wireless networks (LAWNs) are evolving from simple connectivity layers into intelligent fabrics populated by goal-driven aerial agents. In this emerging landscape, unmanned aerial vehicles (UAVs) must autonomously navigate complex tradeoffs between mission-critical objectives, such as timely delivery, and opportunistic utility, such as sensing data harvesting. However, existing control paradigms face a dilemma: model-based optimization is often computationally prohibitive and overly conservative, while black-box DRL offers limited interpretability and does not readily provide the operational assurance required for regulation-sensitive airspace. This article proposes a structured agentic intelligence framework that bridges this gap. By embedding the analytical structure of optimal control into a learnable neural architecture, we develop a grey-box policy with interpretable threshold adaptation and structured execution logic. Our case study of time-critical urban logistics shows that the proposed structured approach can adapt effectively to stochastic environments and maintain a high mission success rate under the considered simulation settings. A favorable tradeoff is demonstrated between delivery reliability, sensing utility, and energy efficiency, while supporting lightweight inference in the tested edge-hardware setting.