The rapid rise of artificial intelligence has put compute power in the spotlight, yet the connectivity layer that ties AI systems together is often overlooked. As organizations move into the agentic AI era—where autonomous agents request data, trigger actions, and collaborate across distributed multi-cloud environments—the limitations of traditional networks become glaringly apparent. The networks that served the VoIP and internet era were never designed to handle the unpredictable, real-time, and always-on traffic patterns that AI agents generate. To thrive, enterprises must rethink and modernize their IP infrastructure, or risk being left behind by more agile competitors.
The New Traffic Reality
For decades, network engineering revolved around the concept of the “busy hour”—a predictable peak period when traffic volumes spiked, typically during business hours. Capacity planning was straightforward: overprovision for the peak, and monitor for anomalies. AI breaks that model. AI agents operate around the clock, making decisions in microseconds and exchanging data continuously. Their interactions are not limited to human working hours; they run late-night data processing, support global operations, and adapt to changing conditions in real time. This creates an always-on traffic profile with sustained, fluctuating demand that is impossible to forecast using traditional models.
Legacy networks were built for voice, video, and general internet browsing—traffic that is relatively predictable and tolerant to slight delays. AI workloads, on the other hand, demand low latency, high throughput, and instant path adjustment. When an AI agent needs to retrieve a dataset, collaborate with another agent, or trigger a financial transaction, the network must respond in milliseconds. Any delay can lead to degraded model performance, missed business opportunities, or even safety issues in critical sectors like healthcare. The old approach of static routing and manual intervention is no longer sufficient.
The Need for Real-Time Telemetry
One of the first gaps operators encounter is the lack of real-time visibility. Traditional network monitoring relies on periodic static reports that may be minutes or hours old. In an AI-driven environment, conditions change in seconds, making those reports obsolete by the time they reach the engineer. Real-time telemetry is now essential, providing operators with a continuous stream of data on latency, jitter, packet loss, and bandwidth utilization. This information enables automated systems to detect anomalies and intervene instantly, rerouting traffic before a bottleneck affects critical AI applications.
Without such telemetry, operators are forced into reactive manual troubleshooting—a slow, costly process that hampers AI innovation. With real-time data, network management becomes proactive. Operators can identify trend shifts, predict congestion, and adjust policies dynamically. This is not just a nice-to-have; it is a prerequisite for running AI workloads with reliable performance.
Moving Beyond Legacy IP Architectures
To meet the demands of agentic AI, organizations must evolve from bloated, rigid IP architectures to more modern designs based on segment routing and Ethernet VPN (EVPN). These technologies provide a foundation for network convergence and precise path control, enabling traffic to be routed dynamically as AI agents’ connectivity needs change.
In the past, network architects had weeks to make changes to support new applications. They could manually configure tunnels, adjust routing policies, and test new connectivity paths. Today, network conditions must change within seconds to keep up with AI agents. Traditional IP networks and legacy protocols like RSVP-TE were designed for a slower, more predictable era. They are too complex and too static for the dynamic demands of AI. Segment routing, by contrast, leverages existing network investments while simplifying operations. It encodes the path a packet should take directly into the packet header, eliminating the need for complex signaling protocols and stateful tunnels. This makes it easier to reroute traffic, adapt to failures, and optimize for specific performance objectives.
EVPN complements segment routing by providing a scalable, flexible foundation for Layer 2 and Layer 3 services. It enables seamless connectivity across distributed data centers and multi-cloud environments, which is exactly where AI agents operate. Together, segment routing and EVPN give operators the agility to create virtual networks tailored to specific workloads, without the overhead of traditional VPN technologies.
FlexAlgo: Flexible Path Optimization
Another critical capability is FlexAlgo, short for “flexible algorithm.” This feature allows the network to calculate optimal paths for different types of traffic based on user-defined constraints. For example, one class of traffic might be optimized for low latency, another for maximum available bandwidth, another for resiliency, and another to satisfy data sovereignty requirements. Network operators can define performance objectives and constraints, and the network automatically computes and maintains the appropriate paths.
FlexAlgo delivers 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, which was difficult to scale and maintain. FlexAlgo simplifies this by moving intelligence into the routing protocol. It enables a single physical network to support multiple logical topologies, each optimized for different SLAs. As AI agents with varying requirements share the same infrastructure, FlexAlgo ensures that traffic is matched to performance requirements rather than constrained by static, one-size-fits-all rules.
For instance, a healthcare organization running an AI diagnostic tool might require absolute latency guarantees, while a financial trading agent might prioritize bandwidth for large data bursts. FlexAlgo allows the network to treat these flows differently, ensuring that each agent gets the resources it needs without impacting others. This level of granular control is essential as enterprises increasingly support a mix of AI and traditional workloads.
Security: MACsec as a Foundation
Modernizing the network is not just about performance; security is equally critical. AI agents often handle sensitive data, whether it is patient records, financial transactions, or proprietary business intelligence. Ethernet-based encryption using MACsec (Media Access Control security) provides link-layer encryption that protects data as it travels across the network. Unlike upper-layer encryption, which can introduce latency and complexity, MACsec operates at the physical layer, providing high-throughput, low-latency security without sacrificing performance.
Integrating MACsec into the IP network architecture ensures that data remains confidential and tamper-proof, even when traversing third-party or managed network services. For enterprises in regulated industries, this is a key enabler for complying with data protection regulations and internal security policies.
Real-World Implementation in Healthcare and Finance
Recently, a group of large enterprises in critical sectors such as healthcare and finance has begun incorporating these capabilities into their network architectures. These organizations need to support a mix of AI and traditional workloads while ensuring that traffic adheres to strict policy, sovereignty, and SLA requirements. By combining real-time telemetry, segment routing, EVPN, FlexAlgo, and MACsec, they can build a network that automatically enforces business policies and performance objectives.
One operational model these enterprises use is deploying and managing their own IP networks over leased optical services from providers. This gives them direct control over their network architecture and the ability to customize policies for specific AI applications. Alternatively, they can consume the same capabilities through a fully managed network service, outsourcing the complexity to a service provider. This creates new opportunities for providers to deliver differentiated, value-added services that go beyond simple connectivity.
Regardless of the operating model, the result is the same: a network that prevents connectivity bottlenecks, maintains service assurance, and scales with AI adoption. Instead of reacting to problems, the network anticipates and adapts. Instead of relying on best-effort routing, it optimizes for each workload. Instead of being a bottleneck, connectivity becomes an enabler of AI-driven innovation.
The Opportunity and the Risk
AI presents both an opportunity and a challenge for service providers and large enterprises. On the one hand, those who modernize their IP networks can monetize the next wave of AI services. They can offer premium connectivity options, guaranteed SLAs, and value-added services such as managed network security and real-time analytics. On the other hand, those who remain tied to legacy architectures risk being overtaken by competitors who embrace network evolution. The pace of AI innovation will not wait for slow, manual network changes. It demands a network that is as dynamic and intelligent as the agents it supports.
The organizations that succeed will be those that view the network not as static infrastructure, but as a strategic asset that must evolve in step with AI. They will invest in real-time telemetry, adopt modern routing protocols, and build security into the foundation. They will move away from busy-hour thinking and embrace always-on, workload-aware connectivity. The time to start is now—because in the agentic AI era, network readiness is business readiness.
Source: Network World News