TLDR: MX-AI is a pioneering system that uses AI agents, specifically Large Language Models (LLMs), to manage and control 5G Open Radio Access Networks (RANs) in real-time. It allows network operators to interact with the network using natural language for both monitoring (observability) and making changes (control), demonstrating performance comparable to human experts with low latency. The system integrates a multi-agent architecture with a live 5G testbed, showcasing a significant step towards AI-native 6G networks.
The future of mobile networks, specifically the upcoming Sixth Generation (6G), is envisioned to be deeply integrated with Artificial Intelligence (AI). This means that networks will be observed, understood, and reconfigured by autonomous AI agents working together across different parts of the network, from cloud data centers to the very edge of the network.
A new system called MX-AI has been introduced as the first complete AI-driven platform designed for this purpose. It connects to a live 5G Open Radio Access Network (RAN) testbed, uses a network of AI agents powered by Large Language Models (LLMs) within the network’s Service Management & Orchestration (SMO) layer, and allows for both monitoring (observability) and control of 6G RAN resources using simple, natural language commands.
The research paper highlights that MX-AI performs exceptionally well, achieving a mean answer quality of 4.1 out of 5.0 and 100% accuracy in decision actions when tested with 50 realistic operational queries. It also boasts a low end-to-end latency of just 8.8 seconds when using advanced LLMs like GPT-4.1. This performance suggests that MX-AI can compete with human experts, proving its practical use in real-world scenarios. To foster further research, the creators have made the agent network, prompts, and evaluation tools publicly available. You can see a live demonstration of MX-AI in action by visiting the YouTube link provided in the paper: MX-AI Research Paper.
The Need for AI-Native Networks
Current mobile networks struggle to meet the demanding requirements of 6G, which aims to connect a trillion devices, offer sub-millisecond latency, ensure extremely high reliability, and operate within strict energy budgets. Traditional, static radio-resource managers are simply not agile enough. Instead, every part of the network – cells, slices, and user connections – needs to be managed by a responsive, data-driven control system that can sense, reason, and act faster than human operators.
The Open RAN (O-RAN) Alliance has addressed this by separating RAN control into different time loops: real-time, near-real-time, and non-real-time. This allows third-party intelligence to be plugged into each loop. While open-source solutions like OpenAirInterface (OAI) and lightweight controllers like FlexRIC provide the necessary hooks for manipulating the air interface, the question of what intelligence should occupy these hooks remained open. Traditional AI methods like reinforcement learning are good for specific tasks but struggle to generalize across the diverse goals of 6G.
This is where LLMs come in. They have shown a remarkable ability to process various types of data, understand human instructions, and break down complex problems into actionable plans – a skill set perfectly suited for network operations. While earlier prototypes existed, they were often limited to simulations or single-agent systems without closed-loop control. MX-AI changes this by introducing an end-to-end system where a network of cooperating LLM-based agents manages a live 5G Open and AI-native RAN.
How MX-AI Works: A Multi-Agent System
MX-AI operates within the Service Management & Orchestration (SMO) layer of the network, specifically at the R1 interface. This placement allows it to operate at non-real-time intervals (1-12 seconds latency), which is still fast enough to compete with human experts and allows for complex reasoning.
The system uses a multi-agent graph, which includes:
-
An **Orchestrator Agent** that interprets operator requests, breaks them into smaller tasks, and activates other agents.
-
Specialized Agents like the **Monitor**, **Planner**, **Policy Synthesizer**, **Validator**, and **Executor**, all connected in a structured way.
These agents receive live performance data, alarms, and network topology information. They refine their understanding using techniques like Retrieval-Augmented Generation (RAG) and by using various tools. Finally, the Executor agent either provides answers to monitoring questions (like KPI summaries or root-cause analyses) or implements control actions (like network blueprint management or policy changes) directly to the RAN.
Real-World Applications: Observability and Control
MX-AI focuses on two main use cases:
-
**Observability:** Operators can ask questions about the network in natural language, such as “How close is slice #2 to its latency SLA over the last hour?” The Monitoring Agent retrieves the relevant data, and the LLM summarizes it, even generating charts for auditing.
-
**Control:** The system can dynamically manage network blueprints and apply policies. For example, a command like “Increase VPN slice reliability to 99.999% until 18:00” is processed by the Deployment Agent, which then creates an updated policy to instruct the network to prioritize that slice.
The system continuously updates its understanding of the network state through a novel push-based mechanism, where changes in network resources and application states are streamed directly into the agent’s knowledge base. This ensures the agents always reason with the most current information, which was found to be crucial for high-quality observability.
Also Read:
- Advancing 6G Network Management with Generative AI: A Hierarchical Learning Approach
- Smart Agents Enhance Mobile Augmented Reality Experience in 6G Networks
Performance and Future Outlook
The evaluation of MX-AI showed impressive results. For control actions, capable LLMs achieved 100% accuracy, meaning they correctly interpreted and executed all commands. Observability was more challenging due to the need to synthesize information from diverse sources, but the system still performed very well. The latency varied depending on the LLM used, with cloud-based models being faster (around 8-9 seconds) and local models (like 70B-class models) offering strong performance with slightly higher latency (12-14 seconds) but providing full data control. Smaller local models achieved very low latency (1.3 seconds) but with reduced coherence.
Crucially, the research found that engineering the data retrieval and tool usage was more impactful on observability quality than simply using larger models. This means that how the agents access and process information is key to their effectiveness.
When comparing the time it takes for an action to be completed (Time-to-Action or TTA), MX-AI agents were competitive with human experts, and in some cases, even faster. This is because the agent automates the process of gathering context, planning, validating, and executing commands, significantly reducing cognitive and search overhead for operators.
While MX-AI represents a significant leap towards AI-native 6G, challenges remain, including further reducing latency for emergency procedures, ensuring agent safety and alignment, developing open agent ecosystems, and scaling to the massive data volumes expected in 6G. However, the current results indicate that agentic control at the R1 interface can already match expert operators and has the potential to surpass human speed and reliability for common network workflows by a significant margin.
