The technology already monitors a wide array of networking information, such as configuration data, log messages from devices and monitoring data. AI triggers alerts when networks behave anomalously and offers reasons for the abnormal behavior. It also creates step-by-step plans to fix problems once they occur.
What is agentic AI and how does it work?
Agentic AI is a type of AI that uses software systems known as agents to enable systems to make decisions autonomously and act independently of human intervention. AI agents are designed to exhibit goal-directed behavior. In the context of the network, AI agents work to keep the network functioning at expected levels and maintain network configuration according to company security policies.
In addition, agentic AI can show some level of environmental awareness, such as knowing not to restart a switch as part of routine maintenance during business hours. Like their non-agentic counterparts, agentic AI systems can create multistep plans and adapt plans to changing circumstances. But AI agents can execute those plans as well as more broadly pursue policy and behavioral objectives with minimal human intervention.
Agentic AI in networking
Agentic AI constantly cycles through the four stages of the OODA loop -- observe, orient, decide and act -- and continues to learn as it goes. Agentic AI operates as follows:
Observe. Identifies what happens in the network.
Orient. Analyzes and understands the data based on its past learning.
Decide. Determines which actions it should take in response based on the data.
Act. Executes its decisions.
Agentic AI requires other tools to act in the network. Although autonomous, agentic AI needs to interact with other components of a system, such as devices, software or APIs, to build upon their capabilities. Examples of ways agentic AI can interact within a network include the following:
Writing a program or script that works through APIs directly on a physical or virtual network appliance.
Sending requests with APIs to other management tools to accomplish tasks, such as asking the management console for a vendor's network gear to push out a software patch immediately.
Using an agent-to-agent protocol, such as Model Context Protocol, to enable a different AI agent to take action.
In some cases, AI agents must interact with network teams to perform actions. Sometimes AI agents can't perform the required actions because they don't have access to the necessary tools. Other times, they require human intervention, such as when swapping out a failing network interface on a router or replacing a dead switch.
Network teams should deploy agentic AI capabilities with as much autonomy as they feel comfortable permitting.
Agentic AI use cases and cautions
AI agents are likely to take on moderately complex network operational roles. Examples of agentic AI use cases include the following:
Identify network problems, diagnose causes and rectify the problem.
Make policy-based changes to the network's configuration when policy changes.
Administrators can determine the level of autonomy AI agents have within the network. Organizations that want to maintain control can program AI agents to have little control and only operate in certain, well-defined situations. For organizations that want more efficiency, network administrators can permit AI agents to run in a broader, but still limited, range or perform independently on anything they are permitted to access.
Still, agentic AI remains a new concept. Network teams should deploy agentic AI capabilities with as much autonomy as they feel comfortable permitting. Most organizations are cautiously assessing how to engage with these tools. Network teams have long memories of automation tools that failed to live up to their expectations and, in some cases, caused significant harm when given too much latitude.
As a result, most companies initially design agentic AI policies with a human in the loop. In these cases, the agent must seek permission to implement planned changes. Once agentic AI shows itself to be capable and reliable, organizational staff will be comfortable enough to assign more responsibility and freedom.
John Burke is CTO and a research analyst at Nemertes Research. Burke joined Nemertes in 2005 with nearly two decades of technology experience. He has worked at all levels of IT, including as an end-user support specialist, programmer, system administrator, database specialist, network administrator, network architect and systems architect.
