Intent-Driven Agentic AI for Safe Energy-Efffcient Open RAN Optimization via a 5G Network-in-a-Box Digital Twin

As networks evolve with virtualization and disaggregation, operators face increasing complexity, rising energy costs, heterogeneous devices, and highly dynamic demand. This work responds to the challenge of enabling networks operating with high levels of autonomy, understanding the directives of the operator in natural language, reasoning over scenarios, and acting intelligently to optimize themselves using Agentic AI. This paper presents an end-to-end demonstration architecture that couples a portable 5G network-in-a-box with an AI-enabled Open RAN (O-RAN) digital twin for intent-driven energy-efficient network optimization. The framework is deployed in a cloud-native configuration aligned with O-RAN principles, exploiting Amazon Web Services (AWS) scalable rApp hosting, elastic compute, and MLOps support, while the VIAVI AI Radio System Generator (AI-RSG) which provides a realistic digital twin for repeatable benchmarking and safe policy validation. The framework contains two tightly coupled agentic subsystems: (i) an intent and scenario generation agent that translates natural-language operator inputs into structured simulation schemas; and (ii) a traffic forecasting and energy optimization agent based on a Long Short-Term Memory (LSTM) model, which provides short-horizon per-RU load prediction and confidence-aware three-class classification to guide base station sleep decisions. A purpose-built energy-efficiency optimizer executes these control actions on a real OpenAirInterface gNB testbed with Universal Software Radio Peripheral (USRP) hardware and commercial user equipment (UE), where controlled channel impairments provide labeled ground-truth events for supervised learning. Experimental results demonstrate a 45.6% reduction in RAN power consumption while maintaining full UE connectivity, corresponding to 6.03 kWh/day energy savings (513 kg CO2/year). The LSTM forecasting model achieves 96.3% training accuracy and 95.7% validation accuracy, with stable validation loss, indicating strong generalization across traffic regimes. These results demonstrate a practical pathway toward safe and autonomous energy management in next-generation 5G/6G RAN systems.