As organizations move from experimental generative AI projects to production-grade agentic AI systems, the conversation is naturally dominated by compute, data and model quality. Yet one of the most critical enablers of the AI era is frequently treated as an afterthought: the network. An AI agent is only as useful as the data it can access quickly and the actions it can trigger in real time. If the underlying network is unable to provide the required connectivity, performance and policy controls, even the most powerful AI infrastructure will disappoint.
This realization is creating a new dynamic for IT and network teams. Those that modernize their IP networks are unlocking new revenue streams from AI-driven services, because they can support the rigorous performance and security demands that modern AI workloads require. Those that postpone transformation and continue to operate aging, static networks run the risk of becoming irrelevant in an environment where speed and agility determine market position.
The new traffic reality of agentic AI
Agentic AI represents a major departure from earlier enterprise workloads. Traditional applications generated traffic based on predictable patterns, with peak periods and quiet hours. Today’s AI agents are continuously active. They monitor streams, consult knowledge bases, invoke APIs, exchange tokens with other agents and initiate workflows across distributed clouds and data centers. They can run at any hour, often with an always-on profile that makes the old concept of network busy hours obsolete.
The workload pattern is also more sensitive. A single action may require that an agent retrieve a record from an on-premises database, call a cloud-hosted inference service, combine the result with another agent’s output and deliver a response in milliseconds. Each dependency creates a potential bottleneck. The network must therefore provide more than high bandwidth; it must offer dynamic path selection, strict latency controls and automated policy enforcement. Otherwise, the distributed nature of agentic AI creates an unpredictable web of north-south and east-west traffic that a traditionally engineered network cannot handle.
Why traditional networks fall short
Legacy IP networks were built for an era of voice, video and internet browsing. These applications could tolerate some latency and could be supported with circuitous routing and manual traffic engineering. Network architects typically had weeks to implement changes, and periodic adjustments based on static reports were enough to keep services running.
That operational model collapses under AI workloads. AI agents interact and tackle tasks at machine speed. A delayed route update or a broken traffic policy can result in lost transactions or, worse, non-compliance with strict data sovereignty rules. Real-time telemetry is essential to enable operators to understand traffic patterns and trigger automated actions. Without it, network teams are forced into reactive manual troubleshooting, relying on reports that may be outdated by the time they are opened.
The same applies to the architecture itself. Many networks have become more complex over time as VLANs, VPNs, access control lists and routing protocols have been layered on top of one another. Rigid architectures make it hard to introduce new AI services or to scale them beyond a limited pilot. There is a clear need for a modernized IP foundation that uses fewer, more capable protocols and supports automated path control.
Segment routing and EVPN as the new foundation
A practical starting point for network evolution is the move away from complex, rigid IP architectures toward segment routing and Ethernet VPN.
Segment routing works by encoding the path a packet should follow through the network in the packet header itself. It removes the need for many per-flow signaling protocols and gives the source node complete control over the route. This capability is especially useful for AI workloads because network operators can steer traffic away from congestion, onto low-latency links or through specific geographic paths. Because segment routing operates with existing MPLS or IPv6 infrastructure, it does not require a costly, disruptive network replacement. Instead, it provides a clean evolution path from current architectures while bringing the control and automation required by AI.
EVPN complements segment routing by providing flexible, standards-based virtual networking for data centers and wide-area networks. It enables consistent policy and seamless workload mobility across locations, which is a necessity for agentic AI that often operates across multi-cloud and hybrid environments. Together, segment routing and EVPN create a converged network foundation that supports both traditional and AI workloads.
FlexAlgo for intelligent path selection
Adding segment routing and EVPN is not enough by itself. Networks also need to calculate the optimal path for different traffic types automatically. That is what FlexAlgo, short for flexible algorithm, enables. With FlexAlgo, operators can define separate algorithms for latency, bandwidth, resiliency, or even data sovereignty and then instruct the network to route different classes of traffic accordingly.
A network operator in healthcare might use one FlexAlgo instance to route medical imaging traffic on a guaranteed-bandwidth path, and another to route patient data along a path that remains within national boundaries. A financial services provider could use a low-latency algorithm for trading orders and a high-resiliency path for settlement traffic. These algorithms replace manual tunnel engineering and static policies with automatic, network-wide path computation. The benefits resemble those once pursued with RSVP-TE, but without the accompanying complexity, per-tunnel state and management burden.
In an AI-heavy environment, the number of traffic classes and SLAs continues to grow. FlexAlgo makes it possible to enforce a wide range of quality objectives on a single physical network. It also enables operators to quickly adjust performance parameters as changing business conditions require, a key attribute for supporting dynamic agentic workflows.
Security and policy in an automated network
Network automation does not reduce the importance of security; it makes security more configurable and more consistent. MACsec, for example, is now being integrated into modern IP networks to provide authenticated, encrypted communications at layer 2. This protects data moving between routers, switches and other network equipment from interception or tampering. For any AI deployment handling personal data or proprietary algorithms, such protection is essential.
Modern networking also allows security policies to be tied to application traffic automatically. As an AI agent connects to cloud services and on-premises resources, the network can classify the traffic, place it on the appropriate path and apply the correct set of access controls and inspection policies. This ensures that compliance is not dependent on individual administrators or on static rules that become obsolete in a rapidly changing application environment. It is particularly important in highly regulated industries, where data sovereignty and residency requirements require that certain traffic never crosses certain borders.
Real-world adoption in regulated industries
Forward-looking network teams are already incorporating these capabilities into their digital transformation strategies. In healthcare, for example, hospitals and research centers are under pressure to use AI for diagnostics, patient monitoring and operational efficiency. Their networks must carry imaging files, electronic health records, real-time telemetry and agent-driven workflows while remaining HIPAA-compliant and secure. Segment routing, FlexAlgo and MACsec together allow them to treat each traffic class based on its own requirements, without maintaining multiple parallel networks.
In finance, AI is increasingly used for fraud detection, customer service, risk analysis and algorithmic trading. These use cases demand low latency and precise traffic controls. A modern IP network with segment routing and FlexAlgo can meet these requirements while still supporting the many conventional applications that a bank runs. It can even facilitate new service offerings for providers. Depending on operational preferences, enterprises may build and manage such networks themselves over leased optical circuits, or they may buy them as a fully managed network service. The second model is particularly attractive as service providers look for new differentiators, creating an ecosystem of intelligent, policy-aware network services.
The strategic imperative of network evolution
Every enterprise needs to decide how quickly it will evolve its network. Agentic AI is not a temporary trend; it is becoming the primary interface for automation and data-driven decisions. At the same time, networks are not the seemingly neutral infrastructure they once were. They determine which applications can scale, which innovations can be delivered reliably and which markets a business can serve.
Modernizing IP networking is therefore a strategic decision, not merely a technology refresh. Organizations that recognize the critical role of connectivity will build the runway for AI innovation. They will be able to introduce new services in days, enforce service levels programmatically and establish trust with customers and regulators through built-in security and compliance.
The winners in the agentic AI era will be those that have cleared the network path before the demand arrives. The integration of real-time telemetry, segment routing, EVPN, FlexAlgo and MACsec security creates a networking foundation that can support AI agents wherever they operate, while preserving the stability required for traditional services. Those that take a proactive approach will be able to navigate the next wave of digital innovation instead of being constrained by the limitations of legacy infrastructure.
Source: Network World News