Agentic-AI Framework for Integrated Design, Implementation, Testing, and Operation of Digital Twin Networks

Abstract

Digital Twin Networks (DTNs) are emerging as a key enabler for the robust optimization and lifecycle management of next-generation networks. However, current DTN approaches still rely heavily on manual design, implementation, testing, and validation of the DTN system, which are often tailored for specific application domains and lack generalization. This time-consuming creation and adaptation is incompatible with the highly dynamic nature of sixth-generation (6G) networks. Consequently, the DTN must be capable of adapting to the dynamic network environment, necessitating automated DTN generation and operation. Crucially, no unified framework currently supports an agentic workflow covering the entire lifecycle, from use-case inception and requirement specification to operation. This paper addresses this gap by presenting an adaptable, agentic framework to jointly design, implement, test, and operate DTNs in a closed loop. The proposed architecture is designed to be extensible, such that it can be reconfigured for other network domains (e.g., radio access, mobility) via its adaptable configuration mechanisms (e.g., agent prompt engineering, comprehensive data collection and control interfaces). However, both the current implementation and experimental evaluation are scoped to core network management and topology optimization. Subsequently, the framework autonomously orchestrates continuous perception, reflection, planning, code generation, testing, validation, and control. We evaluate the framework's capability through two distinct 6G use cases: Multi-Access Edge Computing (MEC) service migration and network slicing. Extensive evaluations with state-of-the-art Large Language Models (LLMs) demonstrate the framework's ability to reliably translate high-level intents into valid actuations that fulfill the specified Key Performance Indicators (KPIs), achieving high success rates while revealing distinct model-specific engineering strategies.