Leveraging AI Agents for Autonomous Networks: A Reference Architecture and Empirical Studies

The evolution toward Level 4 (L4) Autonomous Networks (AN) represents a strategic inflection point in telecommunications, where networks must transcend reactive automation to achieve genuine cognitive capabilities–fulfilling TM Forum’s vision of self-configuring, self-healing, and self-optimizing systems that deliver zero-wait, zero-touch, and zero-fault services. This work bridges the gap between architectural theory and operational reality by implementing Joseph Sifakis’s AN Agent reference architecture in a functional cognitive system, deploying coordinated proactive-reactive runtimes driven by hybrid knowledge representation. Through an empirical case study of a Radio Access Network (RAN) Link Adaptation (LA) Agent, we validate this framework’s transformative potential: demonstrating sub-10 ms real-time control in 5G NR sub-6 GHz while achieving 4% higher downlink throughput than Outer Loop Link Adaptation (OLLA) algorithms and 85% Block Error Rate (BLER) reduction for ultra-reliable services through dynamic Modulation and Coding Scheme (MCS) optimization. These improvements confirm the architecture’s viability in overcoming traditional autonomy barriers and advancing critical L4-enabling capabilities toward next-generation objectives.

💡 Research Summary

The paper addresses the critical transition of telecommunications networks toward Level 4 (L4) Autonomous Networks (AN), where true cognitive capabilities—self‑configuration, self‑healing, and self‑optimisation—are required to deliver “zero‑wait, zero‑touch, zero‑fault” services. While current AI‑driven automation largely stalls at conditional autonomy (L3) due to what the authors term “intelligence plateau” and “acceleration resistance,” this work bridges theory and practice by implementing Joseph Sifakis’s dual‑driven AN Agent reference architecture in a real‑world 5G Radio Access Network (RAN) Link Adaptation (LA) scenario.

The architecture separates a proactive subsystem (System 2‑like deliberation) from a reactive subsystem (System 1‑like fast response) and coordinates them through a central Workflow Coordinator Runtime. Long‑Term Memory (LTM) is realized as a hybrid knowledge store: a Neo4j graph database holds structured 3GPP specifications, while a FAISS vector database keeps unstructured telemetry embeddings. Retrieval‑Augmented Generation (RAG) and symbolic reasoning fuse these sources, providing both logical consistency and adaptability.

Functional modules are built on a tiered technology stack. Situation awareness combines Kalman filtering with LSTM‑based trend prediction; self‑awareness leverages large language models (LLMs) via few‑shot prompting to translate high‑level triggers (e.g., operator commands, anomalies) into abstract meta‑goals. Decision making employs a lightweight multilayer perceptron for initial candidate ranking, deep reinforcement learning (DRL) for multi‑objective optimisation, rule‑based safety checks, and Monte‑Carlo Tree Search (MCTS) for final action planning. The reactive runtime processes sensory inputs, updates LTM, builds predictive models, generates candidate goals, evaluates them, and executes validated actions—all within a closed‑loop feedback cycle. The proactive runtime continuously monitors internal state against predefined thresholds, generates meta‑goals when deviations exceed limits, assesses feasibility, and feeds the same predictive‑planning pipeline to achieve long‑term optimisation.

The authors instantiate this framework in a RAN LA agent that must select the appropriate Modulation and Coding Scheme (MCS) every millisecond based on channel quality indicators. Experiments on a 5G NR sub‑6 GHz testbed demonstrate:

• Average control latency of 9.8 ms, satisfying the sub‑10 ms requirement for real‑time link adaptation.
• A 4 % increase in downlink throughput for enhanced Mobile Broadband (eMBB) compared with the conventional Outer Loop Link Adaptation (OLLA) algorithm.
• An 85 % reduction in Block Error Rate (BLER) for Ultra‑Reliable Low‑Latency Communication (URLLC), bringing error rates down to the 0.1 % range.

These gains stem from the hybrid proactive‑reactive approach: the proactive subsystem anticipates future channel conditions via learned meta‑goals, while the reactive subsystem reacts instantly to sudden SINR fluctuations. The system also exhibits self‑evolution; it continuously refines its policy through online learning, adapting to non‑stationary channel dynamics without human intervention.

Key contributions are twofold. First, the paper delivers the first documented implementation of Sifakis’s AN Agent architecture in a telecom‑grade environment, providing a concrete roadmap for achieving L4 autonomy. Second, it validates a comprehensive AI stack—graph‑vector hybrid knowledge bases, LLM‑driven goal abstraction, DRL‑guided multi‑objective selection, and safety‑ensuring rule engines—showing measurable performance improvements over legacy methods.

The authors conclude by outlining future work: extending the architecture to multi‑agent coordination across network slices, integrating security and privacy safeguards, and evaluating the framework on emerging 6G high‑frequency bands. Overall, the study demonstrates that a rigorously engineered, dual‑driven AI agent can overcome traditional autonomy barriers and bring cognitive, self‑governing capabilities to next‑generation mobile networks.