You’ve tried an AI chatbot for troubleshooting, possibly for scripting. It helped—generally. However your Monday nonetheless begins the identical method: manually constructing lab topologies, writing configurations from reminiscence, and documenting adjustments that no person reads till one thing breaks at 2 a.m.
The issue isn’t that AI doesn’t work. It’s that almost all community engineers are nonetheless on the primary two rungs of the potential ladder.
Three ranges of AI for community engineering
Stage 1: Conversational AI. You ask an LLM to “generate a BGP EVPN configuration for my leaf switches,” and it offers a generic response—it doesn’t know your naming conventions, addressing scheme, or validated design patterns. Helpful for brainstorming, however the mannequin has no entry to your surroundings.
Stage 2: AI Assistants. The LLM features entry to exterior sources—documentation by way of RAG, APIs, information. Cisco’s AI Assistant in Catalyst Middle—powered by the Deep Community Mannequin—is an effective instance: it queries your community state and provides context-aware solutions. However for a multi-step workflow like constructing a lab topology, you’re nonetheless prompting one motion at a time.
Stage 3: Agentic Frameworks. A single or multi-agent AI structure takes your necessities and orchestrates an entire multi-step workflow—utilizing instruments, area data, and your workforce’s requirements—with you reviewing at essential steps. You outline the “what.” The agent handles the “how.”
The leap from Stage 2 to Stage 3 isn’t about smarter fashions. It’s a couple of completely different structure.
What makes an agentic framework
4 core elements make this work for community engineering:
The AI agent is the reasoning engine—an LLM that interprets necessities, reads abilities, calls instruments, and decides the subsequent step. In superior setups, a number of brokers collaborate—a planning agent designs the topology whereas a validation agent checks the output.
Abilities are markdown information that encode your workforce’s area data—naming conventions, design patterns, templates. When a senior engineer leaves, their experience leaves with them. Abilities seize it in a format brokers eat immediately—runbooks the AI really follows.
MCP (Mannequin Context Protocol) servers bridge brokers and your infrastructure APIs—Catalyst Middle, vManage, CML, ISE—to learn state, push configurations, or validate adjustments. As a result of MCP is an open normal, the identical servers work throughout any suitable framework.
Human-in-the-loop gates are obligatory pause factors the place the agent waits on your approval. Nothing touches your infrastructure with out express sign-off. This isn’t a limitation—it’s the characteristic that makes enterprise adoption doable.
What this appears like in apply
Contemplate a typical activity: constructing a BGP EVPN cloth lab in Cisco Modeling Labs for a buyer proof-of-concept.
Handbook: 2-4 hours. Incomplete documentation. Data stays in a single engineer’s head.
Agentic Framework: 10-Quarter-hour. Full documentation generated. Requirements utilized each time.
Engineer request to “Construct a BGP EVPN cloth — 2 spines, 2 leaves, OSPF underlay, iBGP overlay with VXLAN.”
Agent generates a plan — lab identify, 6 nodes, 8 hyperlinks, base configurations, boot order. Presents it for evaluate.
Engineer evaluations, adjusts the VXLAN VNI vary, approves.
Agent executes by way of MCP — create_lab → add_node (×6) → add_link (×8) → set_node_config → start_lab.
Agent verifies all nodes are energetic, BGP EVPN neighbors established, VXLAN tunnels up. Generates documentation.
The agent isn’t producing textual content — it’s executing a workflow. It reads talent information on your requirements, calls MCP instruments to work together with the CML API, pauses on your approval, and produces reusable artifacts.
Constructing your first agentic workflow
You could have the framework—brokers, abilities, MCP servers, human gates. Now you want a workflow: a selected automated course of like constructing a lab or validating a design. Agentic frameworks like Claude Code, OpenCode, Windsurf, and Cursor all help MCP and may orchestrate these workflows. The instance repository makes use of Claude Code to stroll via the total sample:
Outline abilities—Markdown information that seize your workforce’s area data. The repo consists of ready-to-use abilities for EVPN cloth requirements, naming conventions, and IOS XE configuration templates. Begin with one workflow you repeat weekly and encode the choices you make each time.
Join MCP servers—every server bridges an agent to a selected platform API. The repo features a CML MCP server you’ll be able to level at your lab occasion. CML is the best place to begin: low threat, excessive repetition.
Configure brokers—outline what every agent does and the way they collaborate. The repo features a planning agent that generates topology designs and a validation agent that checks the output. You evaluate and approve between steps.
Create instructions—chain the workflow right into a single invocation: parse necessities → generate plan → human gate → execute → validate → doc.
When requirements change, you replace one talent file, not retrain an individual. Each agent interplay advantages from it.
Clone the repo, level the MCP server at your CML occasion, and run your first agent-assisted EVPN cloth construct in beneath half-hour.
The shift that issues
This isn’t about changing community engineers—it’s concerning the emergence of the AI-augmented community engineer. AI doesn’t simply velocity up execution. It reshapes how engineers design, troubleshoot, doc, and protect data. Specialised brokers can plan topologies, validate configurations, or troubleshoot points in parallel—compressing hours of labor into minutes. Talent information codify years of tribal data that will in any other case stroll out the door when a senior engineer leaves. The engineer’s position shifts from activity executor to orchestrator, curator, and decision-maker.
That shift calls for guardrails. LLMs hallucinate—they will generate plausible-looking configurations with fallacious subnet masks or nonexistent CLI instructions. Human-in-the-loop gates aren’t non-compulsory—they’re a core architectural requirement that retains the engineer in management as AI takes on extra of the workflow.
Cisco is already transferring on this path—Meraki’s Agentic Workflows, AgenticOps, and the Deep Community Mannequin all embed AI throughout community operations. The method described right here is complementary for engineers who want customized workflows or multi-platform orchestration.
The deeper impression is organizational. Agentic frameworks flip particular person experience into shared functionality. Design patterns turn out to be abilities. Validated designs turn out to be templates. Data that takes months of onboarding to switch turns into accessible on day one—and improves with each interplay.
Begin small. Choose one workflow you repeat each week. Construct one talent file. Encode what you already know. Run your first agentic workflow construct. The shift from chatting with AI to working with an AI agent is smaller than you suppose—and the impression is bigger than you anticipate.
