Intent-Driven Autonomous Reconfiguration for Telecom OSS: A Generative-AI-Empowered Hierarchical Multi-Agent Architecture

Telecom Operation Support Systems (OSS) still depend on domain experts and ad-hoc scripts for network and service reconfiguration, a process that is labor-intensive, errorprone, and fragile whenever cross-domain dependencies-such as allocating an IP address before tearing down a VLAN—must be respected. The heterogeneity and dynamism anticipated for 6Gclass networks only magnify the shortcomings of today’s rigid, rule-based automation pipelines. This paper introduces an intent-driven, fully autonomous reconfiguration framework that integrates Generative AI with a hierarchical Multi-Agent architecture. At its core, a ReWOOenhanced Chief Agent transforms natural-language intents into a dependency-aware execution graph, utilizing a formally defined Command Pattern to explicitly specify actions and associated data for unambiguous inter-agent coordination. Atomic tasks-annotated with domain-specific constraints and keys governed by a Primary Key Pattern-are delegated to ReAct-based Specific Agents responsible for IP management, network control, cloud orchestration, and service management. Evaluation on a four-domain OSS testbed exposing over thirty RESTful APIs confirms accurate intent parsing, robust key-value consistency enforcement, strict adherence to create-before-destroy constraints, and seamless scalability from single-domain provisioning to complex cross-domain orchestration without code changes. The proposed work delivers a reusable agent blueprint decoupling global planning from domain execution; formal interagent interfaces curbing hallucination and enabling modular growth; an execution-aware extension to ReWOO for structured action planning; domain-driven guidelines for encapsulating OSS subsystems into loosely coupled agents; and a public OSS testbed for benchmarking future Agentic-AI architectures-collectively advancing OSS reconfiguration toward scalable, maintainable, and 6G-ready network operations.