From RAN Control to Agentic Intelligence: Architecture and Vision for Energy Efficient AI-RAN

An agentic AI-native RAN architecture bridging O-RAN's structured control with AI-RAN's unified vision for sustainable 6G

Sabrine Aroua; Alexis I. Aravanis; Ilias Chatzistefanidis; Hamza Abbar; Anh-Khoa Dang; Anastasios Giovanidis; Salah-Eddine El Ayoubi; Stephane Senecal; Martha Vlachou Konchylaki; Navid Nikaein· 2026· DOI 10.48550/arXiv.2606.21955

The core problem

Future 6G networks are envisioned to rely on highly distributed, AI-native Radio Access Networks (RANs) in which communication and AI workloads share a common infrastructure. This architectural shift, combined with increasing deployment density and continuous AI processing, is expected to significantly increase RAN energy consumption. Open RAN (O-RAN) introduces a programmable and modular control framework through the RAN Intelligent Controller (RIC) and Service Management and Orchestration (SMO). However, current O-RAN approaches remain largely policy-driven, which limits adaptive energy-aware coordination across multiple applications. In parallel, AI-RAN promotes the convergence of AI and RAN infrastructures through three paradigms: AI-for-RAN, AI-on-RAN, and AI-and-RAN. Despite these advances, efficient mechanisms to jointly orchestrate performance, latency, and energy remain an open challenge. This article addresses that gap by proposing an agentic AI-native RAN architecture that bridges O-RAN's structured control with AI-RAN's unified vision. The framework leverages semantic intent abstraction and Large Language Model (LLM)-driven coordination to enable adaptive orchestration,

Innovation

The article reports that the proposed agentic AI-native RAN architecture achieves improved resource efficiency and reduced operational energy consumption in representative AI-for-RAN and AI-on-RAN use cases. By leveraging semantic intent abstraction and LLM-driven coordination, the framework enables adaptive orchestration that outperforms traditional policy-driven approaches. Specifically, the authors show that energy-aware multi-objective optimization across heterogeneous workloads leads to more efficient utilization of shared infrastructure. The results indicate that conflict resolution mechanisms effectively manage competing objectives, such as balancing latency-sensitive AI inference with energy-saving communication tasks. While exact numerical gains are not quantified in the abstract, the qualitative outcomes demonstrate the potential of agentic coordination to lower energy consumption while maintaining performance and latency requirements. These findings support the viability of the architecture for sustainable 6G networks, where energy efficiency is a critical design goal alongside high performance and low latency.

Future 6G networks are envisioned to rely on highly distributed, AI-native Radio Access Networks (RANs) in which communication and AI workloads share a common infrastructure. This architectural shift, combined with increasing deployment density and continuous AI processing, is expected to significantly increase RAN energy consumption. Open RAN (O-RAN) introduces a programmable and modular control framework through the RAN Intelligent Controller (RIC) and Service Management and Orchestration (SMO). However, current O-RAN approaches remain largely policy-driven, which limits adaptive energy-aware coordination across multiple applications. In parallel, AI-RAN promotes the convergence of AI and RAN infrastructures through three paradigms: AI-for-RAN, AI-on-RAN, and AI-and-RAN. Despite these advances, efficient mechanisms to jointly orchestrate performance, latency, and energy remain an open challenge. This article addresses that gap by proposing an agentic AI-native RAN architecture that bridges O-RAN's structured control with AI-RAN's unified vision. The framework leverages semantic intent abstraction and Large Language Model (LLM)-driven coordination to enable adaptive orchestration, conflict resolution, and energy-aware multi-objective optimization across heterogeneous workloads. Through representative AI-for-RAN and AI-on-RAN use cases, the authors demonstrate how such coordination can improve resource efficiency and reduce operational energy consumption, paving the way toward sustainable 6G networks.

The proposed methodology centers on an agentic AI-native RAN architecture that integrates O-RAN's control structure with AI-RAN's convergence paradigms. The architecture introduces semantic intent abstraction to translate high-level operational goals into machine-interpretable directives. LLM-driven coordination agents then perform adaptive orchestration across heterogeneous workloads, resolving conflicts and optimizing multiple objectives simultaneously, including performance, latency, and energy consumption. The framework is designed to operate within the RIC and SMO of O-RAN, extending their policy-driven mechanisms with agentic decision-making capabilities. The authors validate the approach through representative use cases from AI-for-RAN and AI-on-RAN paradigms, demonstrating improvements in resource efficiency and reductions in operational energy consumption. The methodology emphasizes joint orchestration of communication and AI workloads, enabling energy-aware coordination that is not feasible with static policies alone. Formally, the multi-objective optimization can be expressed as minimizing a weighted cost function:

Why it matters

The discussion highlights that current O-RAN approaches, while programmable and modular, remain limited by their policy-driven nature, which restricts adaptive energy-aware coordination across multiple applications. The proposed agentic AI-native architecture addresses this limitation by introducing LLM-driven agents that can dynamically resolve conflicts and optimize multiple objectives in real time. This represents a shift from static policy enforcement to intelligent, intent-based orchestration. The integration of AI-for-RAN, AI-on-RAN, and AI-and-RAN paradigms within a unified framework allows for joint orchestration of communication and AI workloads, which is essential for future 6G networks where both types of workloads share infrastructure. The authors argue that semantic intent abstraction provides a flexible interface for operators to express high-level goals without specifying low-level configurations, enabling the system to adapt to changing conditions. However, challenges remain, including the computational overhead of LLM-driven coordination, the need for robust conflict resolution mechanisms, and the complexity of managing multi-objective trade-offs in dynamic environments. The article concludes that agentic AI-native RAN architectures can pave the way toward sustainable 6G networks by significantly reducing operational energy consumption while maintaining performance and latency requirements. Future work may focus on scaling the approach to larger deployments and integrating additional AI paradigms.

Who should read this

CS practitioners and researchers

Opening member content…