• Introduce a new agentic paradigm- symbiotic agents -that pairs large language models with optimizers and examine it through the lens of trustworthy AI. • Implement two concrete agent designs for dynamic RAN control and multi-tenant SLA negotiations. • Leveraged open-source real-world platforms, including OpenAirInterface (OAI) for the 5G user equipments (UEs), RAN and Core Network, and FlexRIC for the Radio Intelligent Controller (RIC). • Utilized real-world datasets with RAN channel quality fluctuations from moving vehicles to emulate realistic mobility scenarios. • Proved that optimization algorithms maximize the decision accuracy of LLM agents with up to 5 times lower error. • Evaluated both larger and smaller models (SLMs) proving that SLMs retain sufficient near-RT performance (82 ms loop) with substantially smaller GPU overhead (99.9 % less). • Introduced and validated a novel LLM-driven architecture towards AGI networks leveraging the developed agents. Large Language Model (LLM)-based autonomous agents are expected to play a vital role in the evolution of 6G networks, by empowering real-time decision-making related to management and service provisioning to end-users. This shift facilitates the transition from a specialized intelligence approach, where artificial intelligence (AI) algorithms handle isolated tasks, to artificial general intelligence (AGI)-driven networks, where agents possess broader reasoning capabilities and can manage diverse network functions. In this paper, we introduce a novel agentic paradigm that combines LLMs with real-time optimization algorithms towards Trustworthy AI, defined as symbiotic agents . Optimizers at the LLM’s input-level provide bounded uncertainty steering for numerically precise tasks, whereas output-level optimizers supervised by the LLM enable adaptive real-time control. We design and implement two novel agent types including: (i) Radio Access Network (RAN) optimizers, and (ii) multi-agent negotiators for Service-Level Agreements (SLAs). We further propose an end-to-end architecture for AGI-driven networks and evaluate it on a 5G testbed capturing channel fluctuations from moving vehicles. Results show that symbiotic agents reduce decision errors fivefold compared to standalone LLM-based agents, while smaller language models (SLM) achieve similar accuracy with a 99.9 % reduction in Graphical Processing Unit (GPU) resource overhead and in near-real-time (near-RT) loops of 82 m s . A multi-agent demonstration for collaborative RAN on the real-world testbed highlights significant flexibility in service-level agreement and resource allocation, reducing RAN over-utilization by approximately 44 %. Drawing on our findings and open-source implementations, we introduce the symbiotic paradigm as the foundation for next-generation, AGI-driven networks-systems designed to remain adaptable, efficient, and trustworthy even as LLMs advance. A live demo is presented here https://www.youtube.com/watch?v=WQv61z1deXs&ab\_channel=BubbleRAN

Beyond 5G and future next-generation networks will have to cope with the ever-growing traffic demand for mobile traffic, as well as low-latency communications. Network densification has been long proposed as a solution for augmenting the available wireless links with more technologies, thus enhancing the available capacity for the end-users. Nevertheless, selecting the optimal split of traffic among the available links is not a trivial decision. Machine Learning (ML) approaches can assist in these decisions, by forecasting metrics collected directly from the RAN, towards predicting the near-future performance, and appropriately selecting the split of traffic. In this work, we evaluate a total of 22 different ML models in such a traffic steering use case, towards determining the solution that yields the best results in terms of accuracy of predictions, training time, and computational resources. We use a real-world testbed prototype based on OpenAirInterface to evaluate our contributions, and use realistic mobility datasets for emulating client mobility. Our results show that the different algorithms can present variations in terms of the achievable throughput, but several can substantially improve the offered wireless network capacity.

As networks become denser and more heterogeneous different paths can be considered in order to reach each multi-homed UE, offering optimal performance. 5G and beyond networks feature contributions related to the dynamic programming of the network, from the operator side, in order to optimally allocate resources in the network. In this work, we consider such a case, where network access is provided to the end-users via heterogeneous (3GPP and non-3GPP) Distributed Units (DUs), converging to a single Central Unit (CU), and programmable on the fly with external interfaces. We employ Machine Learning (ML) methods in order to forecast the Quality of Service (QoS) that a wireless client will get from the network in the near future based on the Channel State Information (CSI) metric. Subsequently, we appropriately steer the traffic over the different heterogeneous DUs for ensuring that the network meets the needs of the UEs. We design, develop, deploy and evaluate our method in a real testbed environment, using emulated mobility. Our results show that the overall throughput of each UE can be drastically improved compared to existing allocation mechanisms.