Intelligent Networks: The Rise of AI Agents in 6G

by

Author: Denis Avetisyan

A new framework leverages large language models to unify control and reasoning across radio access and core networks, promising more adaptable and efficient future networks.

The framework proposes a convergence of Radio Access Network (RAN) and Core Network (CN) intelligence, suggesting a shift towards integrated network control rather than isolated functionality.
The framework proposes a convergence of Radio Access Network (RAN) and Core Network (CN) intelligence, suggesting a shift towards integrated network control rather than isolated functionality.

This review proposes an AI agent-based architecture for RAN-CN convergence, employing the ReAct paradigm to optimize cross-domain performance in 6G.

Despite advances in network automation, current intelligent control systems often struggle with generalization and fragmented decision-making across network domains due to reliance on task-specific models. This paper, ‘Toward E2E Intelligence in 6G Networks: An AI Agent-Based RAN-CN Converged Intelligence Framework’, proposes a novel framework leveraging a Large Language Model integrated with the ReAct paradigm to unify reasoning and control across Radio Access and Core Networks. The resulting AI agent dynamically interprets network data to synthesize adaptive control policies without requiring model retraining, demonstrating enhanced adaptability to unseen scenarios. Could this approach pave the way for truly unified, intelligent control in future 6G networks?

The Illusion of Network Intelligence

Modern mobile networks are experiencing a surge in artificial intelligence applications, yet these deployments often prioritize narrow, task-specific solutions. While AI excels at optimizing individual functions – such as predicting traffic load or managing radio resource allocation – these models typically operate in isolation. This fragmented approach hinders the potential for holistic network optimization, preventing the synergistic benefits that could arise from coordinated decision-making across different network domains. Consequently, opportunities to improve overall performance, enhance user experience, and efficiently manage resources are lost, as the network struggles to adapt to complex and dynamic conditions with the agility of a truly intelligent system. The current paradigm emphasizes specialized AI components rather than a unified, network-wide intelligence capable of addressing challenges from an end-to-end perspective.

Modern mobile networks increasingly deploy artificial intelligence, yet a key limitation arises from the specialization of these models. Each network function – the Radio Access Network (RAN) and the Core Network (CN) – often relies on independently trained, task-specific AI. This creates a form of ‘domain isolation’, preventing seamless information exchange and coordinated decision-making. Consequently, optimization becomes fragmented; for example, the RAN might prioritize immediate signal strength while the CN focuses solely on throughput, leading to conflicting objectives and suboptimal overall performance. This isolated approach hinders the network’s ability to adapt to dynamic conditions and efficiently allocate resources, ultimately impacting the user experience and limiting the potential benefits of AI integration.

Conventional machine learning techniques often falter when applied to the ever-shifting landscape of modern mobile networks, exhibiting limited capacity to generalize beyond the specific conditions of their training data. This necessitates frequent and resource-intensive retraining as network dynamics – user mobility, signal interference, and data traffic – evolve. In contrast, a novel artificial intelligence agent demonstrates robust performance across key network indicators – Reference Signal Received Power (RSRP), data throughput, and location accuracy – achieving results comparable to, and often exceeding, those of established Long Short-Term Memory (LSTM) baselines. This improved generalization capability suggests a path toward more adaptable and efficient network management, reducing the need for constant recalibration and enabling proactive optimization in response to real-time conditions.

Evaluations across radio signal strength, throughput, and location demonstrate that both Seen-LSTM and Unseen-LSTM generalize effectively, achieving comparable performance to large language models.
Evaluations across radio signal strength, throughput, and location demonstrate that both Seen-LSTM and Unseen-LSTM generalize effectively, achieving comparable performance to large language models.

From Silos to Synergy: The Promise of AI Agents

Large Language Models (LLMs) present a viable architecture for integrating reasoning capabilities across the Radio Access Network (RAN) and the Core Network (CN). Traditionally, these network domains operate with separate control planes and limited intercommunication. LLMs, due to their capacity for complex pattern recognition and natural language processing, can ingest and correlate data from both RAN and CN elements. This allows for a unified understanding of network state and performance, moving beyond siloed optimization. Consequently, LLMs facilitate coordinated actions, such as dynamic resource allocation and proactive fault management, that were previously difficult to implement due to the lack of holistic network awareness. This integration shifts network control towards a more intelligent and automated paradigm.

An AI Agent Architecture leverages Large Language Models (LLMs) in conjunction with the ReAct paradigm to move beyond simple predictive outputs to active, iterative problem-solving within a network. The ReAct framework enables the LLM to generate both reasoning traces (thought) and actions, allowing it to observe the network’s response to those actions and refine its subsequent decisions. This contrasts with traditional LLM applications that provide a single response to a prompt; the AI Agent continuously loops through observation, thought, and action, enabling it to address complex, multi-step network challenges that require dynamic adaptation and aren’t solvable with static configurations or pre-defined rules. This iterative process allows the agent to explore the solution space and improve its performance over time without explicit retraining.

A Real-time Monitoring Database (RMDB) serves as the central source of network telemetry for the AI agent, aggregating data from both the Radio Access Network (RAN) and the Core Network (CN). This database incorporates performance metrics, Key Performance Indicators (KPIs), and configuration states, providing a holistic and current view of network status. The RMDB’s architecture is designed for low-latency data ingestion and retrieval, critical for enabling the AI agent to make informed, proactive control decisions. Validated through network slicing implementations, this framework demonstrates improved End-to-End (E2E) optimization by allowing the AI agent to dynamically adjust network resources based on real-time conditions and predicted demands, resulting in enhanced Quality of Service (QoS) and efficient resource utilization.

This AI agent utilizes the ReAct paradigm to iteratively reason and act, enabling a more flexible and adaptive approach to problem-solving.
This AI agent utilizes the ReAct paradigm to iteratively reason and act, enabling a more flexible and adaptive approach to problem-solving.

Automated Policies: The Illusion of Control

Cross-Domain Policy Generation leverages an AI agent to dynamically create network policies based on current network conditions. This process moves beyond pre-defined rules by continuously analyzing real-time network state data – including traffic patterns, device performance, and security threats – to formulate policies optimized for efficiency and responsiveness. The resulting policies are then applied across multiple network domains, enabling automated adjustments to resource allocation, quality of service, and security protocols. This approach has demonstrated a 27.50% SLA Satisfaction (Users) rate, exceeding the 26.63% achieved by combined RAN LLM & CN LLM strategies, indicating improved network performance and user experience.

Automated Network Control leverages dynamically generated policies to directly influence network behavior, minimizing the need for manual configuration and intervention. This is achieved by translating policy directives into actionable commands for network devices, enabling adjustments to traffic routing, quality of service (QoS) parameters, and security settings in response to real-time network conditions. The system’s ability to autonomously implement these changes results in a measurable improvement in Service Level Agreement (SLA) satisfaction, currently reported at 27.50% for users, compared to 26.63% achieved using a combined RAN Large Language Model (LLM) & Core Network (CN) LLM strategy.

Effective operation of the AI agent relies on robust data pipelines that supply current and historical network telemetry, configuration data, and performance metrics. These pipelines must ensure data quality, low latency, and comprehensive coverage of network elements. Furthermore, a multi-agent system architecture is employed to distribute the workload, enhance system responsiveness, and improve resilience against single points of failure. This distributed approach allows for horizontal scalability, enabling the network to adapt to fluctuating demands and maintain consistent performance even under increased load or component failures. The system leverages agent collaboration and redundancy to ensure continued operation and policy enforcement.

An AI-Native Network Architecture is foundational for realizing intelligent network control. Implementation of this architecture demonstrably improves Service Level Agreement (SLA) satisfaction for users, achieving a measured 27.50% satisfaction rate. This represents a performance increase of 0.87 percentage points when contrasted with a combined Radio Access Network (RAN) Large Language Model (LLM) and Core Network (CN) LLM strategy, which yields 26.63% SLA satisfaction. The data indicates that building the network with AI principles embedded from the outset, rather than layering AI on top of existing infrastructure, provides a measurable benefit to user experience as quantified by SLA fulfillment.

The Inevitable Standard: AI as a Necessary Complexity

Mobile network infrastructure is undergoing a significant transformation as standardization bodies like 3GPP and the O-RAN Alliance prioritize the integration of artificial intelligence. This isn’t merely about adding AI as an afterthought; rather, these organizations are actively defining how AI can be natively embedded within the core functions of mobile networks. The goal is to move beyond traditional, manually configured networks towards systems capable of intelligent automation – self-optimization, predictive maintenance, and dynamic resource allocation. By establishing common interfaces and protocols for AI implementation, these standards facilitate interoperability and accelerate the deployment of AI-driven solutions across diverse network environments, ultimately promising more efficient, resilient, and adaptable mobile connectivity.

A crucial component in the integration of artificial intelligence into mobile networks is the deployment of Digital Twins (DTs). These virtual replicas of physical network infrastructure allow for the rigorous testing and refinement of AI-driven policies in a risk-free environment. Before any changes are implemented in the live network, the DT simulates real-world conditions and traffic patterns, enabling engineers to identify potential issues, optimize performance, and ensure seamless integration. This approach significantly reduces the likelihood of disruptions or unintended consequences, fostering confidence in the AI’s decision-making capabilities. By continually validating policies against the DT, network operators can proactively adapt to evolving demands and maintain a consistently high level of service quality, ultimately accelerating the transition to fully automated, intelligent networks.

The integration of artificial intelligence into mobile networks isn’t simply an incremental improvement, but a fundamental shift propelling the development of 6G technology. Unlike previous generations focused primarily on speed and bandwidth, 6G is being designed with inherent intelligence as a core tenet. This means AI algorithms won’t just optimize network performance, but will actively define its capabilities, enabling self-configuring, self-healing, and highly adaptable networks. Future 6G systems envision a seamless convergence of the physical, digital, and human worlds, requiring networks capable of anticipating user needs and dynamically allocating resources – a feat only achievable through pervasive AI. This foundational intelligence promises to unlock entirely new applications, from truly immersive extended reality to real-time remote control of critical infrastructure, marking a decisive leap beyond the capabilities of 5G and establishing a new paradigm for mobile communication.

Traditional network optimization, focused on end-to-end (E2E) performance, often faltered due to the inherent complexities of fragmented network architectures. These limitations hindered comprehensive optimization across the entire data path. However, recent advancements demonstrate a holistic approach consistently surpasses these baseline methods, particularly under dynamic conditions. Studies reveal that by integrating artificial intelligence and employing techniques like digital twins, networks achieve superior performance even when subjected to fluctuating traffic loads-ranging from 10 to 80 users-without the drawbacks of prior, isolated optimization strategies. This represents a significant leap toward more resilient and efficient networks capable of adapting to real-world demands, offering a marked improvement in overall system stability and user experience.

The pursuit of end-to-end intelligence, as outlined in this framework, inevitably introduces another layer of complexity destined to become future tech debt. This convergence of RAN and CN, powered by LLMs and ReAct agents, strives for adaptability, yet history suggests even the most elegant architectures succumb to the realities of production. Grace Hopper famously said, “Humans are allergic to change.” This rings particularly true in network evolution; each leap forward creates new failure modes and unforeseen interactions. The promise of AI-driven optimization is seductive, but the network will always find a way to expose the limitations of even the most sophisticated agent-based system. The core idea of cross-domain optimization simply shifts the problem, it doesn’t eliminate it.

The Road Ahead

This convergence of Large Language Models, AI agents, and the RAN-CN interface presents a superficially elegant solution. The promise of unified, cross-domain optimization is familiar; countless frameworks have chased similar ideals, only to founder on the shoals of real-world deployment. The inherent unpredictability of wireless environments, coupled with the ever-escalating complexity of network protocols, suggests the ‘adaptability’ touted here will require constant recalibration – a new form of maintenance, inevitably. If all tests pass during simulations, it likely means the simulations omit the critical edge cases.

The next phase will inevitably focus on scaling these agents beyond controlled environments. The true test won’t be achieving theoretical efficiency gains, but handling the inevitable collisions when multiple agents attempt to ‘optimize’ the same resources. Resource contention, unexpected interactions, and the sheer volume of data required to train and maintain these models represent significant hurdles. The field will likely see a proliferation of specialized agents, each addressing a narrow slice of the network, before realizing that complexity has merely shifted, not diminished.

Ultimately, the success of this approach hinges not on the sophistication of the AI, but on the robustness of the underlying infrastructure. A beautifully reasoned agent is useless if the network cannot reliably deliver the data it needs, or execute its commands. The pursuit of ‘intelligence’ should not distract from the enduring need for reliable, predictable, and easily diagnosable systems. The elegance of the diagram will, as always, be inversely proportional to the difficulty of debugging it.

Original article: https://arxiv.org/pdf/2602.23623.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-03-02 15:03