Table of Contents
As agentic AI becomes a primary consumer of network capabilities, telecom infrastructure has to evolve from connecting applications to infrastructure that machines can discover, provision, secure and consume
For most of the internet era, the network was built around the fairly stable abstraction that humans use applications, applications use networks, and developers decide how those applications should connect to infrastructure. Agentic AI begins to disrupt that hierarchy.
An AI agent does not request information and wait for a response. It reasons, invokes tools, calls models, accesses databases, communicates with other agents and potentially modifies real world systems. Increasingly, the entity deciding what infrastructure to use, and when and how to use it, could be software operating autonomously on behalf of a person or enterprise.
That makes the AI agent something more consequential than another source of network traffic. It makes the agent a potential consumer of the network itself. The shift is moving quickly. Gartner projected that 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% in 2025. Those agents increasingly move beyond assisting a human user toward completing end-to-end tasks across multiple enterprise systems.
At Mobile Future Forward last week, Qualcomm SVP and GM of Compute and Gaming Kedar Kondap connected this emerging agentic architecture back to the evolution of the smartphone. “Agents are very similar,” Kondap said, pointing to requirements around power, latency and connectivity that the mobile industry has spent decades optimizing. But the fundamental problem becomes more complicated as intelligence distributes across phones, PCs, tablets, glasses, earbuds and other devices, each possessing different information about the same user.
“This hybrid environment of what’s going to be able to run on a device vs. on a cloud” creates an orchestration problem that Kondap called “uncharted territory.”
From distributed applications to distributed intelligence
A conventional distributed application may span multiple infrastructure locations, but its architecture is largely predetermined. Agentic systems introduce another layer of dynamism because the AI itself can determine which model, tool, data source or compute resource it needs to accomplish a task. Cisco SVP/GM of Provider Mobility Masum Mir described the evolution at Mobile Future Forward as AI moving from a GPU and model problem, to a memory and networking problem, and now toward agents interacting with multiple models and other agents. From the service provider perspective, he said AI represented a “demand shock,” and the industry should do more than provision capacity to meet demand — the goal should be to build vertical-specific solutions that deliver tailored outcomes.
AT&T Business SVP Shawn Hakl provided a useful indication of what that looks like inside enterprises. “In the AI world we are seeing on average people using six to 12 models and those models could live anywhere,” he said in an interview at Mobile Future Forward.
The network therefore has to connect more than a user to an application or an enterprise location to a cloud. An agentic workflow may dynamically connect users, agents, models, enterprise data, public clouds, edge compute and other agents. Hakl described AT&T’s architectural response as moving from a physical connectivity layer into a logical layer responsible for functions including policy, security, identity, load balancing and routing, then exposing those capabilities through automated interfaces.
Hakl has described the goal as making the network easily consumed by agents. That gets to the larger architectural shift, but what does an agent-consumable network actually mean?
Ericsson Research argues that telecom is moving from developer experience to “agentic experience.” Instead of a developer manually reading documentation and choosing a network API, an autonomous agent could discover available capabilities, evaluate latency, reliability, price or compliance requirements, and invoke the appropriate service itself. An agent supporting a video application, for instance, might request a quality-of-service capability dynamically based on what it is trying to accomplish.
That means machine-readable APIs are necessary but insufficient. Network capabilities have to become what Ericsson calls machine-delegable, as in the agent needs to understand not only what it can do, but whether it has authority to do it, under what constraints and who bears economic responsibility for the action.
This puts network APIs, identity, authentication, policy and billing into the same architectural conversation. It also potentially changes telecom monetization. Rather than selling connectivity to an application developer that statically configures the network, an operator could expose capabilities that agents autonomously consume as conditions change. But the opportunity is not automatic. As Hakl put it at Mobile Future Forward, “There’s no divine right for us to win in this space. It’s got to get earned.”
Humans click, agents swarm
The network also has to accommodate a radically different operating tempo. Cisco President and Chief Product Officer Jeetu Patel summarized that difference at Cisco Live: “Humans click, but agents swarm.” Agents operate continuously and at machine speed rather than waiting for a human to initiate the next action.
Cisco has attempted to quantify what that means. In one test using an Open Deep Research agent, a single agentic task generated 26.8 MB of network traffic while exploring 44 web sources. Cisco measured a 450% increase in traffic versus performing the task manually, with 70% of the incremental traffic associated with AI inference.
That is one workload, not a universal multiplier, but Cisco’s broader measurements show why agent traffic deserves separate treatment. AI inference flows lasted roughly twice as long as normal web transactions, while about 9% carried more traffic upstream than downstream compared with roughly 0.5% of conventional web flows.
Cisco’s modeling goes considerably further, projecting enterprise network traffic could grow roughly 9x between 2026 and 2035 with agentic AI adoption, versus approximately 2.5x without it. That is a modeled scenario rather than an observed trajectory, and adoption assumptions this far out necessarily carry substantial uncertainty. But the directional point is important: autonomous software can consume infrastructure much faster and more persistently than human-driven applications.
From access control to action control
Traffic is only part of the problem. If an agent can provision connectivity, access enterprise data, invoke a model, spend money or alter a network configuration, knowing that it is authenticated is not enough. The infrastructure needs to understand what the agent is authorized to do.
“As you move into the agentic world, we have to move from just plain access control to action control,” Patel said at Cisco Live. That makes identity, policy, observability and auditability fundamental network functions for the agentic era. Ericsson similarly argues that identity, authentication and validation could become new monetization control points as agents transact on behalf of users and enterprises.
For telecom operators, that may be the more interesting opportunity than simply carrying additional AI traffic. Networks already know something about users, devices, locations, connectivity conditions and service entitlements. They already enforce policy and deliver SLAs across enormous distributed systems. The agentic opportunity is to turn those existing capabilities into services that software can autonomously discover and consume.
That is a materially different proposition from selling more bandwidth. The network of the internet era largely connected people to applications. The network of the agentic era will increasingly connect models, data, tools, compute and machines to one another, and mediate what autonomous systems are allowed to do across them.
The next wave of network users will be armies of agents deciding what they need from the network, requesting it in real time and moving on to the next task before a human would have made the first click.
