Multi-agent AI system uses fine-tuned language models to automate telecom network repairs

Telecom networks have become so large and intricate that maintaining them has grown into a significant operational burden. Network problems currently require Subject Matter Experts (SMEs) to manually sift through multiple data sources, identify root causes, and devise remediation plans—a process that is both time-consuming and resource-intensive.

Researchers have proposed a new approach to this challenge: a multi-agent system that coordinates several specialized AI agents to diagnose and resolve network faults automatically. According to a paper submitted to arXiv, the system leverages fine-tuned smaller language models alongside orchestration workflows to streamline troubleshooting across both radio access and core network infrastructure.

How the system works#

The framework operates through a team of specialized agents that activate once network monitoring tools detect a fault. These agents include:

• An orchestrator that oversees the troubleshooting workflow
• A solution planner that generates remediation strategies
• An executor that carries out fixes
• A data retriever that gathers relevant information
• A root-cause analyzer that identifies what went wrong

The solution planner is a core element. Rather than relying on a large, general-purpose language model, the researchers fine-tuned a smaller language model on internal troubleshooting documentation to ensure the system generates relevant, domain-grounded repair recommendations. This targeted approach reduces the need for vast amounts of labeled training data while improving reliability.

Key advantages#

The multi-agent approach addresses longstanding limitations of AI in telecom management. Traditional models tend to have narrow applicability, require extensive labeled datasets, and often fail to transfer well between different network deployments. By breaking troubleshooting into specialized subtasks and coordinating them through intelligent agents, the system can work across heterogeneous network environments and produce faster, more actionable diagnoses than manual expert review.

Experimental validation shows that the framework significantly accelerates fault resolution in both Radio Access Network (RAN) and Core network domains, suggesting potential for widespread adoption in telecom operations centers.