Agentic AI for RAN Automation: Huawei Proposes Telco-Specific A2A Interface

In September 2025, Huawei hosted the Huawei Wireless Analyst Talk in Shanghai. The focus revolved around how 5G-Advanced enables new services and how Artificial Intelligence (AI) plays a role in the network’s internal and external elements. Unsurprisingly, one of the discussion points revolved around integrating Agentic AI into the Radio Access Network (RAN) to support operator strategies in reaching Level 4 automation for their network.

As part of this discussion, Huawei announced its RAN Intelligent Agent solution, consisting six distinct agents, all of which leverage the Huawei-developed Telecom Foundation Model:

• RAN FME Mate: Supports engineers in troubleshooting and reducing Mean Time to Repair (MTTR) by 30%.
• RAN NOEMate: Utilizes multi-objective decision-making to enable automatic ticket closure and improved quality and efficiency.
• RAN TunningSpirit: Minute-level optimization for dynamic traffic, boosting average user experience by 10%.
• RAN FaultSpirit: Enables proactive Operations and Management (O&M) via multidimensional root cause analysis, reducing fault tickets by 30%.
• RAN WattSpirit: Leverages the RAN Digital Twin System to enable multi-domain energy management while maintaining customer experience.
• RAN AssurSpirit: Leverages the Telecom Foundation Model predictive capabilities to achieve over 90% accuracy for service provisioning.

Huawei adopts a multi-agentic setup with a “leader” agent to manage these domain-specific agents. Introducing Agentic AI into the RAN creates more complexity due to the heightened risk of various synergistic issues. This is why Huawei is also proposing, via The 3rd Generation Partnership Project (3GPP) SA5 and the TM Forum, the Agent2Agent (A2A)-T interface, a telco-specific A2A interface for multi-agent collaboration with increased security and safety compared to the general A2A interface. China Mobile Guandong has experimental deployments of this interface.

The need for AI’s integration within network automation across all domains, not just the RAN, is not a new development and has been ongoing for many years. However, different flavors of AI must be integrated for different use cases. Huawei’s Agentic AI portfolio for the RAN makes it clear that AI agents will be used for the O&M use cases, but also for more real-time optimizations surrounding energy efficiency and service provisioning, which would theoretically include RAN slicing to meet Service-Level Agreements (SLAs).

Huawei is not the only vendor looking to use Generative AI and AI agents for RAN automation. ZTE has also discussed the feasibility of such an interface. Softbank in August announced that it had tested a new Transformer-based AI model for uplink channel interpolation and reported gains of 30% when compared with a baseline model without AI, and 8% when compared with a conventional Convolutional Neural Network (CNN) model in a live wireless environment. The Associated Press (AP) release does say that one of the technical challenges of “AI for RAN” is operating under real-time processing constraints of less than a millisecond. This further explains why Huawei has used Generative AI and Agentic AI more for O&M use cases, which do not typically require such tight constraints, and minute-level responsiveness would suffice. Therefore, it is clear that while these AI flavors have a role to play, the role of classical Machine Learning (ML) will not be going away any time soon.

However, the wider AI industry's concerns about Generative AI and AI agents cannot be ignored for telco-specific models. Hallucinations (whether intrinsic or extrinsic), especially around more high-value use cases such as network optimization at the cell site level, could have profound implications for the network, and multi-agent systems would create new vectors for cyberattacks.

Clearly, the only way for Agentic AI to truly be integrated across the RAN domain globally is through further industry collaboration to standardize the usage and protocols used for safety purposes. However, the industry will face global buy-in challenges, both from a competitive and an operational perspective.

Much of the data in the RAN domain may be structured quite differently telco-by-telco. A global, federated data structure would be necessary to fast-track the implementation of third-party Agentic AI solutions and reduce the need to train/fine-tune the models for each telco, which can be expensive. Though not for commercial settings, we already see this proposed within the AI-RAN Alliance as a Data-for-AI initiative. However, telcos have many legacy systems that rely on already implemented data structures. To change this would be costly and time-consuming, and for competitive reasons, telcos will be extremely resistant to sharing any data. Importantly, O&M data and fault alerts/tickets (especially repetitive faults) are much more structured than other elements in the RAN domain, making Agentic AI implementation easier and requiring less fine-tuning.

Ultimately, incumbent vendors are likely to leverage their existing relationships with telcos, with whom they already have penetration in the RAN automation domain, to develop Agentic AI solutions tailored to their customers and complementing their existing portfolio. The AI-RAN Alliance (of which Huawei is not a member) is unlikely to adopt Huawei’s A2A-T protocol; more likely, it will begin to work on a competing protocol, and within this alliance, we will likely see industry collaboration.