Artificial intelligence has dominated technology planning for several years, but most of the attention has focused on compute: GPUs, accelerators, training clusters, and inference infrastructure. Connectivity has often been treated as a supporting detail. That assumption is becoming dangerous. As AI moves from experimental pilots to production services, and especially as agentic AI begins to operate autonomously, the network becomes a primary determinant of whether AI applications meet their performance, cost, and compliance goals.

Agentic AI describes systems that can plan, reason, and act on behalf of users or other systems. Unlike a simple chatbot that waits for a prompt, an AI agent can request data, call APIs, trigger workflows, and collaborate with other agents across distributed and multi-cloud environments. These agents operate continuously. They do not respect traditional peak-hour models. They create always-on traffic profiles with bursts that can appear in microseconds and spread across data centers, edge locations, and public clouds.

That shift changes the design criteria for enterprise and service provider networks. The old ways of planning for voice, video, and general internet traffic are not sufficient. Organizations need networks that can adapt dynamically as AI agents make decisions and exchange data at machine speed. The question is no longer whether the network can carry more bandwidth. It is whether the network can enforce policy, guarantee latency, maintain sovereignty, and assure service levels while AI workloads scale unpredictably.

Why Agentic AI Breaks Traditional Network Assumptions

Traditional traffic engineering was built around predictable patterns. Networks were sized for busy hours, with capacity planned for the busiest period of the day or week. That model worked well for human-driven applications: email, web browsing, video streaming, and voice calls. It assumed that demand would rise and fall with human activity. It also assumed that changes could be made over days or weeks, because new applications rarely required instant network reconfiguration.

Agentic AI upends those assumptions. AI agents can run around the clock. They can generate traffic that is highly variable, latency-sensitive, and difficult to predict. A single agent may query multiple data sources, invoke models in different clouds, and coordinate with other agents in real time. The network may need to shift paths in seconds, not weeks. Manual troubleshooting and static reports cannot keep up.

Consider a healthcare workflow in which an AI agent reviews patient data, checks for contraindications, and triggers a clinician alert. The agent may need to access electronic health records, imaging systems, pharmacy databases, and external registries. Some of that traffic must stay within a jurisdiction for privacy and sovereignty reasons. Other parts may need ultra-low latency to support real-time decisions. A network that cannot distinguish between these requirements, or that cannot route traffic according to policy, becomes a bottleneck and a compliance risk.

Financial services present similar demands. Fraud detection, algorithmic trading, risk modeling, and customer service agents all rely on rapid access to data and services. A delay of milliseconds can affect outcomes. An outage or misrouted flow can violate regulatory requirements. Networks must provide precise path control, deterministic performance, and strong security without creating operational complexity that overwhelms IT teams.

The Limits of Legacy IP Architectures

Many enterprise and service provider networks still rely on architectures that were designed for earlier eras of VPN and internet connectivity. These networks have served organizations well, but they are often bloated, rigid, and complex. They may use multiple protocol layers, manually engineered tunnels, and extensive state management. They can take weeks to reconfigure for new demands. They may lack the telemetry needed to understand what is happening in real time.

Legacy IP networks also tend to treat all traffic the same, or at least they rely on coarse classes of service. That approach is insufficient when different AI agents and workloads have different requirements. One workload may need the lowest possible latency. Another may need high available bandwidth. Another may need a path that avoids a specific country or region. Another may need maximum resiliency or disjoint paths. Static, one-size-fits-all rules cannot satisfy these varied needs.

Operators often fall back on reactive manual troubleshooting. When performance degrades, they inspect logs, run tests, and try to identify the cause. By the time they have an answer, the traffic pattern may have changed. Static reports are frequently outdated before they are used. This reactive mode is costly and slow, and it cannot support AI services that make decisions in microseconds.

Three Capabilities for the Agentic AI Era

To support agentic AI, networks need at least three foundational capabilities: real-time telemetry for visibility and automation, segment routing and EVPN for convergence and precise path control, and FlexAlgo for traffic-specific optimization. Security features such as MACsec are also becoming essential for protecting traffic in transit and meeting compliance requirements.

Real-Time Telemetry and Automated Intervention

AI workloads demand visibility at a granularity that traditional network monitoring cannot provide. Operators need real-time telemetry to understand traffic patterns, detect anomalies, and support automated intervention. Streaming telemetry and model-driven telemetry can provide continuous data about link utilization, latency, jitter, packet loss, and application performance. When combined with automation and AIOps, this data can trigger closed-loop responses: rerouting traffic, adjusting policies, or scaling capacity before users or agents are affected.

Without real-time telemetry, operators are left with reactive manual troubleshooting. They rely on static reports and after-the-fact analysis. That is not enough when AI agents are making decisions in microseconds and when service-level agreements carry financial and regulatory consequences. Real-time visibility is the foundation for dynamic control. It turns the network from a passive pipe into an active participant in service assurance.

Segment Routing and EVPN for Dynamic Path Control

Modern IP architectures based on segment routing and EVPN provide a more scalable and flexible foundation. Segment routing simplifies traffic engineering by allowing a source node to specify a path as a list of segments. It reduces state in the core, leverages existing network investments, and creates an evolutionary path toward greater agility. EVPN provides a unified control plane for Layer 2 and Layer 3 VPN services, supporting multi-tenancy, mobility, and efficient service delivery.

Together, segment routing and EVPN enable convergence and precise path control. They allow the network to route traffic dynamically as AI agents' connectivity needs change. Instead of waiting weeks to engineer new tunnels, operators can define intent and let the network compute and maintain the appropriate paths. This is essential when AI workloads move between clouds, data centers, and edge locations, and when performance requirements change from one moment to the next.

Segment routing also protects existing investments. It can be deployed incrementally, interoperating with legacy protocols while introducing new capabilities. That matters because most organizations cannot rip and replace their networks overnight. They need an evolutionary path that supports AI workloads while continuing to serve traditional applications.

FlexAlgo for Workload-Specific Optimization

FlexAlgo, short for flexible algorithm, lets the network calculate optimal paths for different traffic types. One class of traffic might be optimized for latency. Another might be optimized for available bandwidth. Another might be optimized for resiliency. Another might be optimized to satisfy data sovereignty requirements. The network can compute paths according to performance objectives and constraints, rather than forcing all traffic through the same static rules.

FlexAlgo delivers many of the traffic-engineering benefits that operators once sought with RSVP-TE, but without the massive complexity. RSVP-TE relied on manually engineered tunnels and extensive state management. It could be powerful, but it was also operationally heavy. FlexAlgo allows operators to define objectives and constraints, then lets the network automatically compute and maintain the appropriate paths. As networks support different SLAs for different AI agents and workloads, FlexAlgo ensures traffic is matched to performance requirements rather than constrained by one-size-fits-all rules.

MACsec and Security by Design

Security is not optional in the agentic AI era. AI agents may handle sensitive data, trigger critical actions, and cross trust boundaries. MACsec provides Layer 2 encryption and integrity protection for Ethernet links, helping organizations protect traffic in transit and meet compliance requirements. For healthcare, finance, government, and other regulated sectors, MACsec can be a practical way to secure high-speed connectivity without sacrificing performance.

How Enterprises Are Putting These Capabilities Together

Large enterprises in critical sectors are already incorporating these capabilities into their network architectures as part of broader digital transformation initiatives. They need to support a mix of AI and traditional workloads while ensuring that traffic adheres to strict policy, sovereignty, and SLA requirements. They may deploy and manage their own IP networks over leased optical services from providers. Alternatively, they may consume the same capabilities through a fully managed network service. Both models create opportunities for providers to deliver differentiated, value-added services.

The result is a network that automatically enforces business policies and performance objectives. It prevents connectivity bottlenecks and maintains service assurance as AI adoption and digital transformation efforts continue to scale. Instead of treating the network as an afterthought, organizations can use it as a control point for policy, performance, and compliance.

For service providers, this shift is both an opportunity and a challenge. AI creates demand for higher bandwidth, lower latency, and more dynamic connectivity. Providers that modernize their IP networks can monetize the next wave of AI services. They can offer differentiated SLAs, secure connectivity, and managed services that help enterprises focus on AI applications rather than network operations. Providers that stand still risk losing relevance as competitors embrace network evolution.

For enterprises, the calculus is similar. Modernizing the network can unlock new revenue from AI-driven services, improve customer experiences, and reduce operational risk. Delaying modernization can lead to performance bottlenecks, compliance gaps, and competitive disadvantage. The network is no longer just a cost center. It is a strategic asset for AI. The organizations that treat it as such will be better positioned to capture the value of agentic AI while managing its complexity. As

Source:Network World News