Network Insight Agent: Enabling AI-Driven Multi-Dimensional Insights and Solution Generation

As a core application of the large model for wireless AI for IT operations (AIOps), the Network Insight Intelligent Agent provides a traffic  stimulation solution for wireless access networks. It delivers in-depth insights into network structure, coverage, capacity, device health, and other dimensions. Leveraging generative AI, the solution generates query strategies, insight summaries, solution recommendations, and multi-dimensional charts—showcasing the powerful capabilities of ZTE’s Nebula Telecom Large Model. The agent supports users in efficiently understanding network insights and solution demands across different stages, application scenarios, and objectives through natural language interaction, enhancing the efficiency of network analysis and solution generation in network planning and optimization.

AI-Driven Applications for Intelligent and Efficient Network Data Analysis and Solution Generation

Traditional network analysis and solution output require operations and maintenance personnel to query large volumes of network data, process it, and provide network planning and optimization recommendations, relying heavily on manual analysis and expert experience. With the introduction of large language models (LLMs), key algorithms and technologies such as intent understanding,  retrieval-augmented generation (RAG), NL2API, NL2SQL, NL2Code, and long short-term memory (LSTM) enhance the intelligence of data analysis and solution generation.

The Network Insight Agent brings the following user values:

• Intent understanding: Leverages the language comprehension and reasoning capabilities of large models to generate network insights through natural language dialogue.
• Report generation: Flexibly generates text summaries, chart presentations, and solution suggestions based on user demands or preferences, creating materials for communication and reporting.
• User adaptation: Adjusts output based on user feedback to better align with their demands.
• Professional enhancement: Through features like knowledge Q&A and guided questioning, users learn while using the system—clarifying concepts, improving problem description skills, and enhancing their problem-solving abilities.

The six key functional features of the Network Insight Agent are as follows:

• Multi-agent collaboration: Multi-dimensional network insight agents, including expert agents for structure, coverage, capacity, and health, collaborate in traffic stimulation scenarios. They provide multi-dimensional insights, offering conclusions and solution recommendations.
• On-demand data query: Provides eight-dimensional insights for wireless access networks, including structure, coverage, traffic, load, health, and  and three additional dimensions (assets, performance, and energy efficiency) to be supported in the future. Users can query statistical data on demand through natural language interaction, enabling a network health check.
• Solution generation: Generative AI, based on expert knowledge bases and insight data, generates insights and solutions.
• Personal agent creation: Users can create personalized combinations of intelligent agents based on their preferences, enabling scene orchestration to cover multiple scenarios.
• Query strategy generation & chart creation: Utilizing the semantic understanding, generation, and logical reasoning capabilities of large models, this feature automates the generation of query strategies, insight conclusions, and solution recommendations.
• Input association: When users input simple terms like “5G” or “coverage,” the Network Insight Agent automatically associates related questions, such as “How many high-load 5G cells are there?” and “What is the distribution of weak coverage cells?” This simplifies user input while demonstrating the product's capabilities.

Multi-Agent Collaboration: In-Depth Application of the Network Insight Agent in Network Planning Scenarios

An intelligent agent provides strong operational capabilities at the core of large models, unlocking their full potential. With clear objectives, the agent can independently think and take actions to achieve these goals. It breaks down the task into detailed steps, utilizing feedback from external sources and its own reasoning to generate prompts to accomplish the goal.

For example, in a network planning application, if a user asks, “What is the network situation around the Datang Everbright City block?” the agent will decompose this task into four main steps: analyzing the sites in the Datang  Everbright City area, determining if the sites are operating stably, checking if the coverage meets the requirements, and evaluating if the network capacity is sufficient.

The multi-dimensional network insight agent first gathers expert agents specializing in network structure, network coverage, network capacity, and device health to analyze the network situation in the area.

• Step 1: The Structure Expert Agent filters sites from the configuration table.
• Step 2: The Health Expert Agent calls relevant interfaces to check for service outages, alarms, and hidden network element issues.
• Step 3 and Step 4: The Coverage Expert Agent and Capacity Expert Agent select the tools for the coverage and capacity insights to execute necessary tasks.
• Final step: The multi-dimensional network insight agent consolidates the information from all expert agents and returns the final response to the user.

To support these user-interactive business processes, the Network Insight Agent is designed using the layered structure shown in Fig. 1.

• Task layer: The agent breaks down the given task and conducts overall planning and task orchestration.
• Execution layer: Based on the task designed, the agent identifies suitable tools and provides the necessary input information to execute the tool’s actions.
• Adaptation layer: Includes session management, prompt management, and external knowledge management, while also providing both short-term and long-term memory capabilities.
• Driver layer: Includes management modules for accessing vector databases, API libraries, and LLM API modules for accessing large models.

The network planning scenarios utilize the multi-dimensional network insight agent and the collaboration of multiple single-dimensional expert agents. As the core agent, the multi-dimensional network insight agent, is capable of task planning, organizing expert analysis, summarizing expert recommendations, and proposing improvement measures. Multiple single-dimensional expert agents query network data in their respective domains to generate solutions.

Personifying the specialized roles of agents and utilizing APIs, knowledge bases, and online data from various dimensions can significantly reduce the hallucination issues of large models and improve the reliability of the generated solutions. Additionally, the collaboration of multiple agents enables problem-solving in multi-goal and complex scenarios.

Currently, the Network Insight Agent, along with ZTE's AIOps system, has been commercially deployed at 24 major sites (provincial-level network management systems) across three major telecom operators in China. During the MWC 2025, held in early March, the Network Insight Agent showcased ZTE's large-model applications on the wireless side and conducted a live interactive demonstration. In the second half of 2025, this application will be implemented in the wireless network management systems of overseas operators, including AIS in Thailand.

The launch of the Network Insight Agent and other agent series marks a new era of intelligence in ZTE's network O&M. Through deep learning and large model technology, it not only addresses inefficiencies and information asymmetry in traditional operation modes but also delivers breakthroughs in multi-dimensional data analysis and expert agent collaboration. Going forward, the Network Insight Agent will evolve to support more precise queries, finer-grained network analysis, and more diverse application scenarios.