AI is rapidly changing how communications service providers (CSPs) optimize the radio access network (RAN), from traffic and capacity management to performance, energy efficiency and service quality. Yet as intelligence matures within individual domains, a larger challenge is coming into view, because the network itself does not operate in domains.
Customer experience and service performance are shaped by interactions between the RAN, transport, the 5G core, cloud infrastructure and the service layer. A problem that surfaces in one domain can originate in another, and the most effective response may require coordinated action across several. I believe the next phase of network autonomy will be decided less by how intelligent each domain becomes than by how well those domains can reason together.
The industry’s progress so far reflects the first half of that equation. TM Forum’s recent assessment finds CSPs increasingly reaching Level 4 autonomy in specific network domains, while autonomy across the network as a whole remains a work in progress. Domain-level autonomy is a genuine achievement, and it also makes the gap between domain decisions and network decisions easier to see.
The Cost of Deciding in Isolation
Consider a familiar degradation. Throughput falls and latency rises across a group of cells during peak hours, and because the evidence available to the RAN points to congestion, traffic is redistributed and radio parameters are adjusted, producing some improvement. The underlying constraint, however, may be a resource limitation in the cloud infrastructure supporting a core network function, in which case the radio intervention treats the symptom and leaves the cause in place. The reverse happens too, when the diagnosis correctly identifies a RAN issue but rerouting traffic in transport or changing policy in the core would resolve it more effectively than another radio adjustment.
Neither case reflects weak engineering within the RAN. Each domain has developed increasingly sophisticated tools, analytics and automation, from RAN intelligent controllers and their radio applications (rApps) to the network data analytics function (NWDAF) in the 5G core. What remains underdeveloped is the reasoning that connects a symptom in one domain to a cause or remedy in another, which today still sits largely between systems, teams and operational processes.
What Shared Context Really Means
Closing that gap starts with shared context, and shared context is more than shared data. CSPs already collect enormous volumes of telemetry, so the challenge lies in turning that information into an operational understanding of how the network behaves as a system.
Three kinds of knowledge make that possible. Topology and dependency information shows which cells, transport paths, network functions and cloud resources support one another and which services rely on them. Real-time conditions show what is actually happening, as opposed to what inventory records suggest should be happening. Historical incidents, remediation outcomes and engineering knowledge provide an operational memory of what has worked before and under which circumstances. Combined, they allow a system to distinguish events that merely coincide from conditions that are causally connected, which is precisely the judgment the degradation above required.
Specialists Working From the Same Picture
Rules-based automation remains effective for predictable scenarios, and it will continue to handle a large share of routine operations. Multi-domain networks, however, produce combinations of conditions that are difficult to encode exhaustively in advance. An agentic system can assemble evidence, form a hypothesis, examine the relevant conditions across domains and determine which action is appropriate, which allows it to work through situations no workflow designer foresaw.
A single general-purpose AI attempting to understand every aspect of the network would struggle to match the depth that radio, core, transport and cloud operations each demand, and the more workable model resembles an experienced engineering team. A radio specialist does not need to become a cloud engineer to contribute to a complex incident, but they do need to recognize when evidence from the cloud or core changes the diagnosis, just as a core specialist may need to understand RAN conditions before recommending a policy change. Specialized agents working from the same shared context can make that exchange happen at operational speed and scale, and when objectives compete, as energy efficiency and peak performance often do, they can weigh them against the outcome the operator has defined for the service.
Continuous feedback makes the model progressively more effective. Actions that resolve problems, recommendations that engineers reject and outcomes that differ from expectations all return to the shared context as operational signals, gradually building a richer understanding of how decisions play out under different network conditions.
Governing Autonomy Across Domains
Cross-domain intelligence also changes the risk profile of autonomy. An incorrect action within one domain usually has a contained impact, whereas an incorrect decision spanning several domains can propagate much further. I would argue that this makes governance part of the operating model itself, designed in from the start rather than added once agents are already acting.
In practice, agents need explicit permissions over what they can access and change, policies that determine which actions may proceed automatically, and escalation paths for situations that exceed established confidence or risk thresholds. Routine, reversible actions with a demonstrated record of reliability can move progressively toward autonomous execution, while higher-impact actions pass through additional verification or human approval. Every action should also be explainable afterward and checked against its intended result once executed, so that trust rests on evidence.
The Changing Role of the Engineer
Governed autonomy changes what human involvement means rather than removing it. As systems become more capable, engineers move from executing routine operational decisions toward setting policies, defining the limits of autonomous action, handling exceptions and deciding when that autonomy should expand. Autonomy stops being a binary state and becomes a scope that CSPs widen deliberately as their systems demonstrate that they can reason reliably and coordinate across domains.
RAN optimization will keep advancing, as will automation across the core, transport and cloud. The operators that pull ahead will be those that pair that progress with shared context and a clear model of delegated authority, so that every improvement in one domain also strengthens the decisions made across the rest of the network.
The views expressed in this article belong solely to the author and do not represent The Fast Mode. While information provided in this post is obtained from sources believed by The Fast Mode to be reliable, The Fast Mode is not liable for any losses or damages arising from any information limitations, changes, inaccuracies, misrepresentations, omissions or errors contained therein. The heading is for ease of reference and shall not be deemed to influence the information presented.