What Is a Network? In the AI Era, the Answer Is Changing.

If you asked a network engineer to define “the network” back in 2016, they’d answer with confidence. Back then, the network was simply the boxes and circuits that connected monolithic applications, including switches, routers, firewalls and MPLS circuits.

Skip forward to 2021, and the answer wasn’t so simple. Many enterprises were using service meshes to support containerized applications, and the definition of the network expanded to include service discovery, sidecars and application-layer policies.

Today, as enterprises deploy distributed AI infrastructure spanning agents, models, data sources, and tools across different locations and environments, the definition of the network is evolving once again. Just as service meshes emerged to connect microservices, new networking constructs are emerging to connect and govern agents.

The network is now enabling agents to discover one another, collaborate securely, access the right resources, and operate within organizational guardrails. To support agentic systems, the network becomes increasingly agentic itself, capable of interpreting intent and adapting its behavior to help achieve business goals.

In my previous blog post, I argued that networks can no longer remain passive; they must become active participants in agentic AI. Instead of just carrying traffic on behalf of agents, networks need to make goal-driven decisions like the agents themselves do. In this post, we’ll explore the defining characteristics of this emerging networking and what they mean for enterprises deploying AI systems.

As the definition of the network has changed over the years, one thing has remained constant: Enterprises don’t actually care about networks for their own sake. What they want is outcomes; the network is just the means to an end. So, before we get into the details of this agentic networking, let’s consider a simple thought exercise: What would enterprises tell their networks if those networks were actually listening?

What enterprises really want to tell their agentic networks

When you strip away the technical details, it’s surprisingly easy for enterprise leaders deploying AI systems to express networking intent in plain business language:

• Sovereignty: “We need every inference workload that touches customer data to stay inside the EU, even during failover. We also need to ensure that data never crosses the public internet.”
• Token optimization: “We need every agent request routed to the lowest-cost model that still meets our quality standards. We also need to cap monthly token spend for each agent.”
• Provenance: “We need full traceability for every agent interaction across our regulated workflows, including where the inference is physically executed.”
• Ecosystem reach: “We need to connect our agents directly to our partners’ agents, model providers, and SaaS platforms. We need privacy and predictable latency for these connections.”

The concept of intent-based networking is not new. It has existed for more than a decade, yet enterprises have struggled to go from concept to reality for two main reasons:

• Networks couldn’t reliably understand business intent.
• Even if they could, network policies were limited to terms like VLANs, QoS and bandwidth.

Enterprises can now bypass both of these barriers. Generative AI has empowered networks to parse natural-language intent. Also, networking vocabulary has expanded beyond traditional constructs to include concepts such as agent identities, models, prompts, and tokens.

The pieces are all available, yet agentic networking remains challenging for many organizations. Let’s look at why that is.

Why do many agentic networking initiatives stall?

With generative AI enabling the front end of intent-based networking, expressing what you want from your network is now the easy part.

A new back end has also emerged. Within the last two years, essentially every classic network function has developed an agentic counterpart:

• Packet routing became prompt routing. Networks can now decide which model, provider or location should handle each request, based on cost, quality and policy.
• Firewalls became guardrails. Policy checkpoints now inspect prompts and tool calls instead of ports and packets. This protects against prompt injection attacks, unauthorized tool use, and other emerging threats.
• Flow monitoring became prompt-level observability. OpenTelemetry conventions have standardized traces of agent invocations, model calls and token usage, much like NetFlow once standardized packet flows.

Agent gateways are now the most prominent building blocks for this new kind of network. They emerged from every corner of the industry, including hyperscalers, established network vendors, open-source projects, and API platforms, and it all happened in the span of two years.

For enterprises, the fact that they have access to agent gateways from many different providers is undoubtedly a good thing. It gives them flexibility to choose the options that best fit their existing networking stack. However, just deploying these agent gateways does not mean they have a true agentic network.

Many agentic AI initiatives continue to stall because there’s a missing middle between the front end and back end: the layer that translates intent into actions across interconnected gateways.

This echoes something we’ve seen throughout history. Whenever new network functions emerge—such as load balancers and firewalls—the functions themselves always come first, and the infrastructure that globally coordinates those functions comes later. This time, the gap is more noticeable because of how quickly the agentic functions emerged.

The semantic control plane fills the missing middle for agentic networking

The layer that sits between the natural-language front end and the functional back end can be defined as the semantic control plane. It translates business intent into coordinated configuration across agentic functions and the established network infrastructure beneath them.

Essentially, the semantic control plane transforms the network into one big AI agent, because it gives the network three common characteristics that all agents share: a goal, the judgement to act on that goal, and the ability to adjust over time in continuous pursuit of that goal.

Some might argue that part of this problem has already been solved, and they have a point: Public model aggregators provide agents with unified routing to hundreds of models from a single API. But intent translation can’t stop at the API routing layer. Aggregators run over the best-effort public internet; therefore, they lack control over network path, workload placement, performance and privacy.

The semantic control plane must address every aspect of intent translation across every dimension, including cost, compliance, performance and observability, all the way down to the physical layer. This is the only way to ensure sovereignty, provenance, token economics and privacy. After all, there’s a big difference between sovereignty that’s declared in a system prompt and sovereignty that’s actually enforced on the wire. The control plane bridges this gap between intent and action.

What enterprise leaders need to know about agentic networking

To evaluate a potential networking solution, whether you’re buying it or building it yourself, ask yourself the following questions:

• Does it accept intent in business terms? You need a solution that incorporates direct natural-language interfaces or MCP—the same agentic interfaces that your business teams already use in other areas.
• Can it translate intent down to the physical layer? API-level routing is not enough. You need a solution that accounts for physical paths, placement, encryption and jurisdiction.
• Is it neutral? Agents need flexibility to access different models and tools across different vendor environments. Therefore, you need a solution that can avoid lock-in and conflicts of interest.
• Can it reach your ecosystem partners? You need direct, private connectivity between your solution and all the model, data, SaaS and partner-agent endpoints you depend on.
• Is it adaptive and capable of self-improvement? You need a solution that can observe outcomes and optimize itself based on changing conditions.

Answering these questions enables you to target your AI investments where they’ll have the biggest impact, instead of over-investing in individual functions. When you apply your budget to agentic networking, you’re driving your AI strategy forward, because you’re ensuring that intent actually compiles throughout your network.

How Equinix approaches agentic networking

At Equinix, we built the new Equinix Fabric Intelligence™ solution specifically to help customers adjust to the changing role of the network in the agentic era. It’s an AI-native control plane that works across three layers:

• Interaction: Chatbot interfaces and third-party agents accept natural-language intent.
• Intelligence: Components include Fabric Super Agent, Fabric MCP Tools and Skills, and Fabric Agent Factory. They work together to translate declared intent into coordinated action across network functions, environments and paths.
• Foundation: Functions run on Equinix’s network of private, neutral interconnection spanning more than 10,000 ecosystem participants across 77 strategic markets. This empowers the intelligence layer to observe and react to changing conditions.

With Equinix Fabric Intelligence, enterprises can ensure their networks are speaking the same language as the business, across every layer and interaction. To learn more, read the white paper: “Why intelligent networks are essential in the AI era.”