Title: Advancing Autonomous O-RAN: A Multi-Scale Agentic AI Architecture for Real-Time Network Orchestration

Abstract: Open Radio Access Networks (O-RAN) offer the potential for adaptable 6G connectivity by leveraging disaggregated, software-defined elements and open interfaces; however, this high degree of programmability significantly heightens operational complexity. Within this ecosystem, various control loops operate simultaneously across the service management layer and the RAN Intelligent Controller (RIC), creating the risk that independently developed control applications may interact in unpredictable or unintended manners. Concurrently, breakthroughs in generative Artificial Intelligence (AI) are facilitating a transition from standalone AI models to agentic AI systems. These advanced systems are capable of interpreting objectives, coordinating multiple models and functions, and dynamically adjusting their behavior over time.

This paper introduces a multi-scale agentic AI framework designed for O-RAN, which structures RAN intelligence into a coordinated hierarchy spanning the Non-Real-Time (Non-RT), Near-Real-Time (Near-RT), and Real-Time (RT) control loops. The framework comprises three distinct agent types: 1. Non-RT RIC: A Large Language Model (LLM) agent that converts operator intent into actionable policies and manages the lifecycle of models. 2. Near-RT RIC: Small Language Model (SLM) agents responsible for low-latency optimization, with the capability to activate, adjust, or deactivate existing control applications. 3. Near the Distributed Unit: Wireless Physical-layer Foundation Model (WPFM) agents that enable rapid inference close to the air interface.

We illustrate how these agents collaborate via standardized O-RAN interfaces and telemetry data. Through a proof-of-concept implementation utilizing open-source models, software, and datasets, we validate the proposed agentic approach in two key scenarios: maintaining robust performance under non-stationary conditions and facilitating intent-driven slice resource control.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC