Understanding the Potential of the Agent-Orientated Network

Machine learning has been used in networks and technology products for many years. But generative AI and agentic AI mark a major change in capability and open new automation possibilities. In telecom networks, agentic AI is most closely associated with the network core and with operations support system and business support system, or OSS and BSS, functions. Agentic AI is just as important in the radio access network (RAN), and there’s enormous activity here.

Agentic AI Solutions Aim to Augment Humans

Conventional AI machine models categorize and inform, but generative AI can add meaning to unstructured data sources and enable agents to act on the information. Agents can perceive conditions, spot patterns across multiple data sources, build plans and execute them. A supervisory agent can manage a series of specialist agents. In essence, this creates a new team that boosts the productivity of an operator’s existing technical teams.

Traditional RAN management relied on fixed algorithms and human intervention to handle network congestion or resolve faults. The introduction of 5G-Advanced has increased network complexity to a point where human-centric operations are no longer sensible or scalable.

A RAN agent isn’t simply a smarter algorithm sitting on top of an existing management platform. It’s a system capable of pursuing a defined objective autonomously over a sequence of decisions and actions. An agent might be tasked with maintaining a target quality of experience across a cluster of cells during a crowded event. To do that, it monitors real-time radio conditions, adjusts beamforming parameters, coordinates handover thresholds with adjacent cells, reallocates spectrum between frequency bands and modifies scheduler weights — iterating continuously as conditions evolve.

Changing Data Traffic Increases the Need for RAN Agents

Network traffic patterns are changing rapidly. This is an area where Ericsson, Nokia and Huawei agree, and all three suppliers discussed the trend at MWC 2026. AI-generated content now accounts for more than 60% of online content, according to Huawei’s analysis presented at the event in March 2026. The interaction patterns associated with AI services — multimodal, conversational, latency-sensitive — differ substantially from the video streaming and web browsing that shaped the design of 4G and early 5G networks.

AI-based network traffic tends to be more bursty and, for visual AI applications, requires a consistently fast uplink with a speed of 20 Mbps everywhere — this is the benchmark all network vendors believe operators must support at a minimum. Huawei goes further and argues symmetrical performance should be the goal. Real-time inference workloads also require consistently low latency and minimal jitter. These are harder requirements to meet than peak downlink throughput, and they need more responsive, fine-grained RAN control than older RAN management tools provide.

There are numerous markets being addressed by agents now. Most of these rely on mobile networks, so the use of agentic AI directly affects mobile network usage. For example, smartphones increasingly include AI agents built-in that can interact through text, voice, pictures and even video. Agents are being deployed in homes to improve device management. Smart cities similarly benefit, because agents can pull together unstructured information. For example, Nokia showed a demo at MWC using generative AI to inform operators about road traffic congestion. Agentic AI is also important for industrial manufacturing and robotics, such as the emerging category of humanoid robotics, like AgiBot. In time, robots may offload some of their compute needs onto AI hardware built into the RAN. Huawei forecasts there will be 100 billion intelligent connections worldwide by 2030.

Into the Agentverse and Deterministic Networks

There are several key areas that Huawei argues are driving the shift to an “agentverse”. End users are no longer just interacting with the wider world through a touch of a button or by writing a few words. Instead they’re speaking to their phone or using the camera on their phone or smart glasses to interpret the world. Such visual AI is also essential for autonomous driving, drones and robotics.

Multimodal AI interaction needs improved uplink performance — as does live streaming — but they have slightly different requirements that the network must accommodate. Huawei reports that multimodal interaction alters the ratio of upload and download data traffic and in tests has seen upload data represent 63% of the total upload and download traffic.

Agents will not work in isolation. They will coordinate in teams. In the RAN, agents will enable better multisite management to improve the overall experience in an area rather than altering settings on each site in turn. Agents will enable a next-generation self-organizing network (SON). They will allow operators to go beyond the limitations of traditional SON designs and reap the full benefits.

With more devices becoming connected, from cars to wearables to sensors, the network has more user complexity. Yet every device, and every type of device, must still have sufficient connectivity to meet its needs. Agents help to make trade-offs where necessary to ensure the best overall connectivity outcome given the network assets and resources that are available. Moving the network from best efforts to a deterministic approach requires the additional versatility that RAN agents provide.

Using an AI-MOS to Make Networks Suitable for AI

In the past, for voice calling or video streaming, the telecom industry has used a mean opinion score (MOS) approach to connect users’ perceptions of network quality with measurable network metrics. The idea is that with a MOS model, an operator’s technical team can know what network quality exists with any given combination of technical characteristics. Then, they can understand how adjusting multiple network configuration settings can improve network quality.

Huawei is now developing a MOS approach to AI too. A significant change is the idea of the Token Block Pipeline. Traditional networks prioritize traffic based on broad quality of service markers. Huawei’s approach uses an AI-MOS to map multimodal demands such as the consistent low-latency needs of an embodied robot to network performance indicators. The Token Block Pipeline enables layered, modality-aware transmission, ensuring deterministic reliability. The AI-MOS model considers four key areas: transmission quality, input quality, decision quality and presentation quality.

Additionally, the Huawei service agent solution is using the standard Agent to Agent Protocol for Telecoms (A2A-T) for inter-agent communication and management. This is a telecom-specific variant of the open A2A standard originally announced by Google. A2A-T is from TM Forum. Huawei uses A2A-T to enable on-demand agent interaction for several types of agent.

High Performance Everywhere Needs All-Band Massive MIMO

Having the agentic tools to manage the RAN isn’t sufficient to deliver on users’ network needs. Nor is it enough to assure an end-to-end network slice with higher quality-of-service levels across a whole network. Operators must fully utilize new spectrum, where available, like the upper-6 GHz band. In most of the world this has been allocated for IMT use. There are a few exceptions like South Korea and the US, which have opted to use it as unlicensed spectrum.

Operators also need the capability in the RAN to serve users in large, dense buildings where higher frequency-capacity band signals don’t reach well. Huawei has unveiled a new 256T massive MIMO active antenna unit (AAU) usable on sub-6 GHz bands, which is reported to support 1 Gbps downlink and 1 Tbps uplink. This unit aims to greatly boost overall capacity and cell edge performance.

With the efficiency of 5G little improved over 4G radio — both use OFDM — massive MIMO has been a key technology to improve performance. It allowed operators to greatly boost capacity on early 5G launches on the 3.5 GHz band. Now, it’s even more important to maximize the benefit from upper-6 GHz bands too. Historically, the challenge with increasing the number of antenna elements has been increased size and weight. This is why it’s notable that newer units benefit from new materials and design that mitigate the problems and enable more elements.

For lower-frequency bands important for in-build, Huawei now has a sub-1 GHz AAU that supports 700 MHz, 800 MHz and 900 MHz bands with 16 TRx. Indicating the strength of Huawei’s wireless portfolio, it again won the Best Mobile Infrastructure GLOMO Award for its ultrawideband (UWB) AAU product line. The range uses a UWB architecture to efficiently aggregate fragmented spectrum resources. It allows the products to increase capacity by a reported four to seven times, improve speeds tenfold or extend coverage by 10 dB compared with less-sophisticated 4T4R equipment.

The UWB AAU can support higher-frequency FDD bands too. For example, in Nigeria it has been deployed in a 1.8 GHz, 2.1 GHz and 2.6 GHz combination to boost traffic volume by a reported 90% on LTE compared with the older 4T4R equipment.

Agents Are Moving into Production

CCS Insight saw many examples of AI agents moving from labs into operator networks at MWC 2026 from many different network suppliers. There were proof-of-concept demonstrations, field trials and increasingly commercial operations. Agents hold enormous promise to allow operators to offer more-sophisticated customer services, flexibly manage network performance and enable differentiated connectivity without increasing headcount.