Performance Assessment of AI-Driven Agents in 5G Networks for Goal-Oriented Communications

Artificial Intelligence (AI)-based agents are expected to play a central role in the evolution toward Sixth Generation (6G) networks, enabling autonomous, adaptive, and goal-oriented network operation. This shift introduces new challenges in terms of reliability, latency, and computational efficiency, particularly when agents interact with critical infrastructures such as core networks. Unlike traditional Large Language Model (LLM) applications, agentic systems operate in dynamic environments, interact with external tools, and exhibit non-deterministic behavior, making their evaluation significantly more complex. This thesis presents the design, implementation, and evaluation of an AI-driven agent operating over a real Fifth Generation Core Network (5G CN) for automated traffic usage monitoring. The proposed proof of concept (PoC) integrates a containerized 5G CN, a real Next-Generation Node B (gNB), tool access based on the Model Context Protocol (MCP), and a LangGraph-based agent. Agent performance is evaluated through an automated pipeline based on the DeepEval framework, using multiple executions and different LLM architectures. The evaluation jointly considers task completion, tool usage, latency, and token consumption. The results reveal a trade-off between reliability, latency, and computational cost, showing that model architecture has a greater impact on agent effectiveness than parameter count alone. No single model consistently outperforms the others; instead, the most suitable choice depends on deployment constraints and performance requirements. Overall, this work demonstrates the feasibility of deploying and systematically evaluating AI-driven agents in real 5G environments, highlighting the importance of multidimensional evaluation frameworks for future AI-native telecommunications networks.