A field guide to agentic AI in networking
Every major networking vendor now ships something called an AI agent. Some of them reason step by step, call tools, and take action on your infrastructure. Others are a chatbot wearing a new name tag. This site tries to tell the difference, and to keep a running account of the open source projects, protocols, and standards work happening underneath the announcements.
What actually changed
Network automation has gone through a few genuine shifts: scripted if/then automation in the 80s and 90s, machine-learning-driven anomaly detection through the 2010s, and generative AI chat interfaces earlier this decade that could explain a problem but not touch it. What's new now is agents that close the loop: they can look at telemetry, form a hypothesis, call a tool to check it, and in some cases make a change, all without a human clicking through each step.
That's useful for the boring, high-volume parts of network operations: chasing down why an access point went offline, correlating an alert across five dashboards, checking whether a proposed ACL change breaks something three hops away. It's also a change in who, or what, has write access to your network, which is why this site spends as much time on the security and governance side as it does on the products themselves.
Read the fuller history and where the industry stands on autonomy →
Explore the directory
Agents and platforms, grouped by what they are rather than how they're marketed.
Full-stack vendor platforms
Marvis, Cisco's AI Assistant, Arista AVA, and the assistants vendors are building directly into router and switch operating systems.
See these agents →Observability & overlay tools
Vendor-agnostic platforms that sit above the network, correlating multi-vendor telemetry and proposing (or executing) changes.
See these agents →Security-centric agents
Agents built for threat detection, policy translation, and SOC triage rather than general operations.
See these agents →Open source & research
MCP servers, community automation projects, and the academic work exploring how agents route tasks between each other.
See these agents →Three things to understand first
The protocol underneath it
Most of this only works because of the Model Context Protocol, and a handful of early IETF drafts trying to bring it to routers and switches directly.
Read about MCP in networking →The security model
Giving a language model tool access to production infrastructure opens a new kind of attack surface. Understand it before you deploy one.
Read about the risks →How far along we are
Vendors talk about full autonomy. Most real deployments are still a few rungs down the ladder. TM Forum's autonomy levels are a useful way to calibrate the claims.
Read about autonomy levels →About this project
NetAgents is a side project, not a product. I maintain it alone, with a background in network automation, and it started as a research pass through the (crowded, fast-moving, sometimes overhyped) landscape of AI agents in networking. If something here is wrong, out of date, or missing, I want to hear about it.