Next-generation 6G radio access networks will need to serve multiple stakeholders, support diverse services, and operate across heterogeneous cloud-edge infrastructures. Conventional network slice controllers remain rigid and disconnected from business context, leaving a persistent service-level agreement (SLA) gap between high-level service objectives and real-time network conditions. This thesis investigates how agentic AI, powered by large language model (LLM) agents, can bring intelligence across the future 6G RAN stack—from business-plane negotiation to near-real-time control—while ensuring fairness, efficiency, and trust.
We first design Maestro, a collaborative business-plane framework in which each operator or vertical is represented by an LLM agent. Stakeholders express high-level intents in natural language, and a mediator agent, assisted by optimisation modules, detects conflicts and guides iterative negotiation to consensus. The resulting service intents are translated into network intents and enforced by RAN slicing functions, aligning business goals with technical feasibility.
To reduce hallucinations and execution overhead, we introduce symbiotic agents, which combine LLM reasoning with deterministic optimisation. We instantiate this paradigm in two cases: (i) an RAN-control agent for dynamic slicing, where the LLM adapts the gain of a proportional controller; and (ii) a multi-agent SLA-negotiation agent, where gradient descent generates Pareto-efficient offers. Experiments on a live OpenAirInterface/FlexRIC testbed show that symbiotic agents reduce decision errors fivefold compared with standalone LLM agents, while properly tuned small language models retain similar accuracy with 99.9% lower GPU overhead. The results show that optimisation grounds LLM decisions and enables near-real-time operation (82 ms).
We then propose Agoran, an agentic marketplace inspired by the ancient Greek agora. Its architecture distributes authority across three autonomous branches: Legislative, which performs compliance checks using retrieval-augmented LLMs; Executive, which maintains real-time situational awareness; and Judicial, which assigns trust scores and arbitrates malicious behaviour. Stakeholder-side negotiation agents and an SRB-side mediator negotiate Pareto-optimal allocations produced by a multi-objective optimiser, reaching consensus in a single round. On a 5G testbed with realistic mobility traces, Agoran improves eMBB throughput by 37%, reduces URLLC latency by 73%, and saves 8.3% of physical resource blocks compared with a static baseline.
Finally, we realise MX-AI, an end-to-end agentic observability and control platform for Open and AI-RAN. MX-AI instruments an OAI/FlexRIC testbed, deploys a graph of cooperating LLM-powered agents in the SMO layer, and exposes RAN observability and control through natural-language intents. On 150 realistic operational queries, MX-AI achieves a mean answer quality of 4.1/5, 100% action accuracy, and 8.8 s end-to-end latency using GPT-4.1, demonstrating the feasibility of AI-native RAN control.
Overall, this thesis establishes a holistic agentic AI framework for 6G RAN automation. It develops LLM-based and symbiotic agents, proposes governance and marketplace architectures, and validates them on live 5G testbeds. These contributions improve network flexibility, enable multi-stakeholder negotiation, support near-real-time control, and strengthen trustworthy, regulation-compliant operation, laying the foundation for intelligent, self-optimising 6G networks.
Overall, this work demonstrates how combining MBRL with FMs leads to more sample-efficient and generalizable decision-making systems.
