Telecom operators face a familiar challenge in modern mobile networks. They collect huge volumes of performance data daily. Yet they still need fast answers. Which cells may degrade soon? Which metric matters most? What action should engineers take?
A new research project, TelcoAgent, aims to answer those questions. It uses AI to forecast 5G network performance. It also links its reasoning to 3GPP standards.
The system appears in a June 2026 arXiv preprint. Its full name is “TelcoAgent: A Scalable 5G Multi-KPM Forecasting With 3GPP-Grounded Explainability.” The project focuses on key performance metrics, or KPMs. These metrics show how network cells behave over time.
TelcoAgent brings three main elements together. It uses a time-series forecasting model for performance prediction. It adds a multi-agent LLM layer for reasoning. It also builds a 3GPP knowledge graph from standards documents.
That graph matters because AI systems can invent confident answers. TelcoAgent tries to limit that risk. It checks its conclusions against standards-defined network functions. In simple terms, the system must show its work.
The research reports zero-shot forecasting across 200 cells. Zero-shot means the model predicts without cell-specific retraining. That could save major time for operators. Training separate models for thousands of sites becomes difficult quickly.
TelcoAgent tested seven KPMs across a three-month U.S. operator dataset. The authors report strong forecasting accuracy against established methods. They also say the system can explain expected degradations. It then suggests operational steps, such as changing parameters.
This approach could help network teams move beyond dashboards. Engineers do not only need warnings. They need context, likely causes, and safe next steps. Standards-based explanations may also support audits and internal approvals.
However, the work remains early. The evaluation covers one operator and 200 cells. That does not prove global reliability. Networks vary by vendor, spectrum band, geography, and traffic behavior.
Deployment also presents a harder question. TelcoAgent recommends actions, but does not prove full automation. Real closed-loop use would require network system integration. It would also need testing with strict operational safeguards.
There is another important balance. 3GPP grounding improves trust and traceability. Yet operators often use local optimization practices. Some may move faster than standards. TelcoAgent must support both compliance and practical flexibility.
The wider industry is moving in the same direction. NWDAF already supports analytics inside 5G systems. Other telco AI benchmarks are also emerging. They aim to measure telecom reasoning, not generic chatbot skills.
For now, TelcoAgent looks like a serious research step. It combines forecasting, explainability, and standards alignment. That combination matters for autonomous network operations. But live production use still needs broader proof.
