Agentic AI for networking: Catalyst or distraction?

Defining agentic AI

First off, what is truly meant by agentic AI? Enterprise Strategy Group's definition helps to explain the essentials of what makes AI agentic, including the following:

• Autonomy. Agents can operate independently.
• Perception. It has the ability to collect, process and interpret data from multiple sources, including other agents.
• Reasoning. It can make informed decisions and analyze past actions to improve future performance.
• Adaptability. It adjusts behaviors based on new information or conditions.
• Memory access. Agentic AI can use data from past interactions, both short and long term.
• Tool use. It can interact with external tools to assist with task completion.
• Goal-oriented. It proactively modifies and optimizes behaviors to improve goal achievement.
• Collaboration. Agentic AI can decompose tasks and engage multiple, complementary, specialized agents to achieve goals.

In my ongoing conversations with enterprise networking and network management technology providers, I've seen that many are on a path to deliver or already deliver agentic AI. However, many vendors use agentic as a marketing term for tools that don't comply with the definition. That doesn't mean they aren't adding value, but it does mean that not everyone means the same thing when they use the term agentic.

I see two camps among those that are not fully agentic, but are attaching to this wave with similar approaches:

1. Agentic-ready. These are products or offerings that do not attempt to deliver a fully agentic system, but can participate in agentic architectures, either by API or, increasingly, by supporting agentic communications protocols such as Model Context Protocol (MCP).
2. Agent-based. These systems have lots of agents, and some of them can have high levels of sophistication, but they typically lack one or more of the characteristics outlined above, such as reasoning and agent-to-agent collaborative capabilities. In most cases, these technologies will evolve to become fully agentic over time.

Agentic AI today

No technology supplier wants to miss this important wave of evolution and value. This has given rise to one of the fastest rates of innovation that the network management tools sector has seen in a long time, and the network equipment manufacturers are all getting into the mix as well.

Without attempting to be exhaustive and complete, here are some illustrative examples from my recent engagements:

• Juniper Networks. We must start with Juniper and its Mist AI-native networking platform because it was really the first mass-market AI system for networking. At the center is the Marvis AI Assistant, and more recently, Juniper introduced its Marvis Minis remote experience monitoring agents. The company recently expanded Marvis Minis support for service-level expectations and client-to-cloud path assessment, more intuitive controls for Marvis Actions, and Marvis Client for gathering and incorporating endpoint telemetry. Given our terminology here, the current iteration of Juniper's platform would best be described as agent-based.
• Extreme Networks. At its recent Connect user conference, Extreme announced limited availability of Extreme Platform One, an integrated management platform that ties together the entire Extreme networking portfolio. Of note here is that this platform uses a wholly new architecture, built from the ground up around agentic AI principles. It covers all aspects of the definition, including problem decomposition, multiple specialized agents that collaborate on tasks, access controls linked back to the human operators that invoke them and more. This might be the first full-fledged agentic AI management platform from a network equipment manufacturer to be delivered.
• LogicMonitor. In 2024, LogicMonitor introduced Edwin AI, an intelligent event correlation assistant that found great favor by reducing alert volumes. This has since evolved into a proper agentic system, announced this spring, with multiple specialized agents including log analysis, change records, metrics analysis and private knowledge base connectors. Also new are direct integrations with OpenAI and ServiceNow. This approach ticks all the boxes of our agentic definition, at least to some degree.
• Kentik. Specialists in hybrid network traffic analysis, Kentik first added AI capabilities in early 2024 via Kentik AI, which included a natural language assistant to facilitate data navigation and Journeys to assist investigative workflows. More recently, the company focused on how to bring its powerful viewpoints into larger AI architectures, likely by adding MCP support on top of existing rich and broad APIs. This is a great example of an agentic-ready approach.
• Itential. Focused heavily on network automation and orchestration, Itential recently launched an MCP server to enable external large language model agents and AIOps platforms to interoperate with the Itential Platform. This is another perfect example of an agentic-ready strategy.
• NetBox Labs. Over the past two years, NetBox has established itself as the new go-to database of record regarding what is deployed in a network and how it is all connected and interrelated. This is a valuable data set for powering all sorts of potential AI functions. Consequently, the company recently announced NetBox Operator, a new design-stage agent project that will use this important data. This follows the release earlier this year of an MCP server for external systems to access NetBox data. Another solid example of an agentic-ready system.

What's next for agentic AI in networking?

Clearly, this is a fast-moving space, and keeping up with the new technologies could easily become a full-time job. Expect more -- a lot more -- over the coming months, as new features and new products are ready to be made public.

Of particular interest, and where we will focus our research efforts, will be the outcomes that agentic AI systems demonstrate to deliver against everyday, real-world tasks and challenges.

For now, networking pros and technology leaders should learn about these systems and watch for three key enablers: good data to feed AI; good structures for managing agents and their authority levels; and appropriate means for integration and orchestration, such as MCP support. If you can check all those boxes, regardless of whether the tech meets the full agentic AI definition, you will be in a good position to make progress with these promising new technologies.

Jim Frey covers networking as principal analyst at Enterprise Strategy Group, now part of Omdia.

Enterprise Strategy Group is part of Omdia. Its analysts have business relationships with technology vendors.