Agentic AI With Zero-Shot Transfer for QoS- and Energy-Aware Resource Auto-Scaling in Cloud-Native B5G Network Slices

Abstract

This work proposes a decentralised multi-agent recurrent soft actor-critic (MARSAC) framework, an agentic AI approach for QoS- and energy-aware resource auto-scaling in cloud-native Beyond-5G (B5G) network slices. MARSAC deploys autonomous agents embedded within network slice environments, each capable of perceiving local states, making policy-driven decisions under uncertainty, and autonomously executing resource scaling actions. The framework addresses joint optimisation of QoS, resource efficiency, and energy efficiency through a novel multi-objective reward formulation balancing slice-specific QoS metrics against system-wide efficiency. Using a multi-agent partially observable Markov decision process with decentralised training, MARSAC achieves linear computational complexity as slices increase, unlike centralised approaches exhibiting exponential growth. MARSAC incorporates domain randomisation to enable zero-shot sim-to-real transfer to unseen real-world traffic conditions without retraining. Experimental evaluations on a real-world testbed demonstrate that MARSAC improves QoS across all heterogeneous slices (eMBB, URLLC, mMTC, V2X). MARSAC achieves an improvement of up to 140% and 77% in resource efficiency over MADDPG and an improvement of up to 20.3% and 56% in energy efficiency over Kubernetes HPA under synthetic Poisson-based and real-world Milan CDR-based traffic, respectively, demonstrating a strong zero-shot generalisation to unseen traffic patterns. Scalability evaluations across 4 to 64 network slices under zero-shot out-of-distribution traffic further demonstrate that MARSAC outperforms MADQN and MADDPG by up to 45.8% and 46.2%, confirming its robustness and suitability for large-scale cloud-native B5G network deployments.