LLM-Driven Multi-Agent Framework for Autonomous Network Slice Management: Architecture and Proof-of-Concept Experiments

Network slicing (NS) enables service providers to create independent virtual networks, storage, and computing resources (so-called slices) from a shared physical infrastructure across one or more operators. These slices can be tailored to specific services, user needs, and Quality-of-Service (QoS) requirements. However, most production-grade orchestration frameworks deployed today still rely on manually crafted templates and centralized controllers, even though a growing body of research has begun to explore deep reinforcement learning (DRL) and multi-agent DRL (MARL) for autonomous and distributed orchestration. As the number of slices increases, limited visibility across administrative domains and human-driven configuration render provisioning slow and error-prone. We address these limitations by proposing a multi-agent management framework where Large Language Models (LLMs) act as intelligent interfaces between tenants and the Management and Orchestration (MANO) stack. Agents interpret high-level intent provided by users or applications, translate it into standards-compliant descriptors, cooperate with other agents and underlying ML-based optimizers to allocate resources, and supervise the slice lifecycle. To validate this paradigm, we conduct two complementary experiments that together cover the AI-driven stages of the proposed workflow. First, we benchmark four state-of-the-art LLMs in generating Yet Another Markup Language (YAML)-based slice descriptions from natural-language prompts of varying complexity. GPT-4.1 consistently produces valid schemas and adheres to constraints, while other models occasionally violate conditions but still yield usable outputs. Second, we implement a proof-of-concept multi-agent system that, from a single user request, models a slice, synthesizes a Topology and Orchestration Specification for Cloud Applications (TOSCA) descriptor, and generates a deployment plan. This second experiment validates the multi-agent transformation chain from a slice request to deployable artifacts. The end-to-end workflow completes in about 40 seconds and produces artifacts validated by a domain expert. The final infrastructure instantiation of these already validated artifacts, e.g., through Kubernetes, Docker Compose, or NFV orchestration tools, is intentionally not treated as a separate AI experiment in this paper. Our results show that modern LLMs, when orchestrated in a collaborative agent framework, can significantly reduce the engineering effort and turnaround time required to move from high-level intents to deployable slice descriptors while ensuring alignment with industry standards. The proposed LLM-based framework can be integrated with any resource allocation optimizers or ML-based (e.g., DRL/MARL) infrastructure and network controllers for multi-tenant resource optimization. We emphasize that the present validation targets the orchestration/specification layer of network slice management, namely the translation of natural-language intents into standards-compliant, internally consistent, and infrastructure-feasible artifacts. Open challenges and future research directions are discussed at the end of the paper.