5
Verizon is moving from scripted RAN automation towards agentic network operations, where AI reasons across domains, shares context and takes action – with the operator retaining control of intent, outcomes and guardrails.
In sum – what to know:
Scripts to agents – Verizon wants to move beyond rule-based automation towards AI systems that can reason about problems and act across RAN, transport, and other network domains.
Fight for context – Verizon argues that vendors can supply powerful RAN agents, but only the operator has the network topology and institutional knowledge to arbitrate between them.
Control and intent – Verizon is working towards more autonomous closed-loop action, but wants humans to define intent, control consequential changes and retain oversight of major incidents.
Maybe it was just the quality of the questions (chapeau, Sean); but Verizon rose above it all at Intelligent RAN Forum last week – above the commercial tests at Orange, the early wins at Optus, the scare stories at SK Telecom, the blueprints at Softbank – to draw the line between automation and autonomy, capability and context, tools and platforms, research and action, order and chaos. It was a useful closing session, which painted the US firm like it had all the right questions, if not the answers just yet; but then probabilistic AI is an awkward bedfellow for a deterministic telco. Which is why Verizon talked at some length about guardrails, arbitration, provenance, and (human) intent.
Verizon wants to move beyond the kind of rule-based automation that drives helpful changes in its own radio (RAN) network – of the sort, actually, that Optus referred to in another session – to an agentic operating model where AI systems reason across its different network domains, and also coordinate with one another, and take actions in response to problems that are not already scripted by humans. Which is the AI dream for Level 4/5 network operations, of course. Verizon’s closed-loop automation platforms processed more than 70 million configuration changes – back in 2025. It has saved however-many thousands of manual labor hours for technicians, it reckons.
Network autonomy
Speaking at the RCR event last week, Anil Guntupalli, senior vice president of technology and product development at Verizon, draws the distinction. “Automation is about what we know, and scripting around it,” he says (a little paraphrased). “Autonomy is about what we don’t know, and reasoning [around it] – and [then] working autonomously within the network and the domain.” The radio network does not exist in isolation, as a silo of self-contained functional headaches – is the point. “It has thousands of KPIs, and dependencies on transport [and] everything else in the environment. So you have to not only predict, but you have to act dynamically in real time… This is where the agentic world kicks in.”
A RAN agent must know whether the problem is really in the RAN, and only in the RAN – and how changes made in response to it by a new model army of ghosts-in-the-machine might cascade across adjacent domains. “The issue could be latency in the backhaul or transport. So how do you communicate that, and put the workflow together?” Sub-domain agents might investigate individual problems, he suggests, before their findings are compiled to determine a “final deterministic action to take”. He says: “That is the challenge we are going against, and working towards.” The agentic network, then, at least in Verizon’s telling, looks less like a collection of clever bots and more like a management system for them.
Which is where mobile operators have always paid their dues, actually – in final orchestration, becoming automated, growing autonomous, of a myriad architecture of network-rooted functions. So a ‘social network of agents’ – to borrow a phrase from Optus – returns multi-domain findings into a top-layer steering platform, which is trained on hard-won carrier knowhow, let rip as rarefied business intelligence. Or something like that. The challenge is to arbitrate between agents – to share context and state from one agent to another, and coordinate actions so they produce a coherent response. “You need someone to orchestrate that very effectively,” he says.
“That’s where we are spending more time.” Which imposes a couple of boundaries, of sorts: between humans and agents, and also between network makers and network operators. The first is not the standard human-in-the-loop story. Besides “physical swap-pouts”, Verizon is happy for agents to make certain changes – configuration and software updates, he says; resetting nodes, dealing with “sleepy cells”. But the engineer defines its remit (“intent”), and therefore its guardrails. “The foundation is the same – the principles in the architecture and intent,” says Guntupalli. In the event of a high-priority (P1) incident, possibly escalated by agents, a proper staffer has to step in.
Managed orchestration
“The human gets the ticket and, ultimately, the truck-roll,” he says, suggesting the agent might dispatch the ticket, and even execute the fix, but that it at least requires human sign-off. “[The agent] cannot alter the state of the network or the experience of the customer. That’s where we [want] the human to [be] – to put some guardrails around it.” It is the marker in the ground for telco AI, as it stands – and where Verizon draws the other boundary line, about its own inherent expertise. “The intent and outcome is something Verizon wants to own,” he says. Its vendor partners bring key innovations with RAN controllers and applications (RIC and rApps), and loaded AI-RAN systems.
“They might bring a small language model, as well. But what they [don’t have]… the context of my entire Verizon – the entire topology… [to] reason across domains.” An energy-optimization rApp might reroute traffic to compensate for a degraded cell without knowing the transport network has no spare capacity. So a clever action might be made into a stupid one, effectively, when the wider context is known. “We are very careful about how the topology works. We are bringing our own intelligence layer on the top… where everything correlates, and the reasoning happens. [That] orchestration… could be proprietary to Verizon, and every operator [will have their own] differentiation.”
Which might be a more interesting battleground in agentic networking: not about which vendor/operator combo has the smartest agent, but which best-owns and -organizes the context for agents to work, operator by operator. “Intent and outcome will be driven by what Verizon wants to do, and how the network needs to work for Verizon,” repeats Guntupalli. “Arbitration is something most people underestimate because context sharing is basically handing off the state of one agent to another. You need someone orchestrating that very effectively.” For that, and to fix agent-made ‘P1’ incidents, they need a kind of digital version of Slow Horses – just to answer Jackson Lambe’s four-effs.
As in: what, where, why, and how – the eff? If an agent changes a network configuration, Verizon wants to know which agent did it, what network state it saw, what data it used, and what context it was operating in. It also wants to keep the number of sub-agents involved sufficiently limited that the chain of reasoning is traceable. That provenance can then feed back into training and improvement. “This is where the differentiation is going to come for the entire industry,” Guntupalli says. “But we don’t want to go back to the pre open-RAN state; otherwise it is going to mean custom integration for each of these agents.” The last point gets into the gnarly question of (necessary) standards.
Industry standards
Open RAN, he notes, is built around deterministic interfaces and specifications; implementations can be tested to behave according to a set of industry-wide specs. But agents are different. Their behavior changes with context; they are required to exchange state and logic as they hand work between themselves. How do you expose a complex and constantly-changing network to an AI agent in a way that is complete, structured, and trustworthy enough for the agent to make decisions? “The short answer [it is] absolutely needs standardization,” he says. Which translates as common approaches to telemetry and agent interaction, potentially extending to “multi-agent reasoning protocols”.
Guntupalli adds: “This is where differentiation will come for the entire industry.” Indeed, beyond clever op-ex gains in automated RAN point solutions, this is the whole game, he says. “Efficiency is interesting, and [the] table stakes – but what’s exciting is what we’re doing every morning when the engineers show up.” Verizon is doing “contextual engineering”, he says – to organize and map data so agents can reason over the top of it, while keeping enough of intelligence close to the network for closed-loop action. There is another aspect, too, about memorializing all of their “institutional knowledge”, accumulated over decades, in digital vaults, to reuse in agents and AR/VR-style gadgetry.
Which frees-up engineers to deal with “harder problems”, he says – such as the somehow-less-glamorous but probably-more-consequential business of building more symmetrical uplink capacity. If sensors, wearables, machines, robots, drones – and whatever else is bundled together as physical AI – increasingly generate information that needs to move upstream, then the old assumption that network traffic comes down the pipe, onto mobile devices, starts to look dated. “Physical AI is about to hit us; all these things need a symmetrical uplink capability,” he says. Verizon is not waiting around for 6G to make the network “AI-native”, he adds.
The telco pyrotechnics that will finally spark are already on the 5G starting blocks, he says: uplink capacity, open RAN, general-purpose compute, multi-vendor networks, integrated sensing (ISAC). He points to a recent ISAC demo with Lockheed Martin and Nvidia, where RAN algorithms were used in 5G spectrum to detect drone noise and flight paths. “We are building that natively, within the network; we don’t have to wait for 6G. [It will] evolve [for] humanoids and wearables, too. We try to make sure the network is built, [and] always working in the background – with agentic native AI working magic in the background… Those are the foundations, and then 6G will roll in with new spectrum and everything else.”
