Empowered, Not Replaced: A Network Engineer’s First Steps into AI

If everything you’re hearing about AI and agents sounds like science fiction, take a step back and take a breath. You’re not behind. The whole landscape shifted in a matter of weeks and months, not years, and the tooling is genuinely new.

How new? We recently published a guide cataloging viable infrastructure servers for Model Context Protocol (MCP), and there are already 60 plus of them. From vendors, from open source, from individual contributors. The protocol itself is less than two years old. So if you feel like you blinked and the ground moved, that’s because it did.

Here’s my message, and it’s the whole spirit of what we’re doing in the VibeOps Lounge: we’re going to go slow, and we’re not leaving anyone behind. This is a journey everyone gets to take. Let me give you the practical first steps.

Start With the Right Fear

A lot of network engineers are nervous about AI. Some fear is fine. Healthy, even. You should not hand an agent the keys to the kingdom on day one. But I want to redirect that fear, because I think most people are worried about the wrong thing.

The thing to actually be afraid of is the day your manager tells you that you have to use AI. It’s coming. Organizations are allocating tokens and tracking token usage. You do not want to be in a one-on-one where leadership asks whether you’re using AI and your answer is that you don’t believe in it or you’re scared of it.

It’s a competitive market to get work right now. AI for network engineers is going to come up in the interview. When it does, you want to be the person who says, sure, I use the NetBox MCP server all the time, what do you want to know?

That’s how you stand out. Everyone has a certification. Everyone has the basics. AI fluency is new enough that you can still be the unicorn.

Treat Your First Agent Like a Junior Engineer

Think about what you do when a junior engineer joins the team. You don’t give them BGP policy changes on day one. You tell them to go learn the network. Go investigate. Go update the documentation. Every one of us has been handed the SSH key and told to fix the docs. It’s a rite of passage, and honestly, it’s a bit of a crappy job.

So what better job for an agent? I am not going to let an AI agent touch BGP policy on day one either. I’m going to start it exactly where I’d start a junior: documentation, testing, compliance, triage, ticket enrichment. There is an enormous amount you can do that is completely safe because it doesn’t change anything.

This is where a decade of muddled thinking trips people up. People have conflated network automation with configuration management for years. They are not the same thing. Yes, you can automate configuration management, but it’s one small slice of what automation actually is.

Read-only is where you start, and it delivers value immediately.

Your Day-One “Hello World”

Let’s get concrete, because I want you to try this. A lot of teams are rightly cautious about sending their configs, their IP addresses, their intellectual property to a cloud model. What if it trains on that data? What if a competitor ends up with access to it? That caution is reasonable. The good news is you don’t have to send your data anywhere.

Open source models have come a long way. There’s still a parity gap between the frontier models from the hyperscalers and the open ones, but that gap is closing fast. So here’s the exercise you can run tonight, air-gapped and private, on your own machine:

1. Download Ollama. Not the model, the framework. It runs locally on your Mac or PC.
2. Pull down a small local model. Google’s Gemma, Microsoft’s Phi, and Mistral are all solid starting points.
3. Turn on the model’s local REST API endpoint.
4. Write a short Python script that pulls your interfaces and asks the model whether they’re healthy.

That’s it. That’s a perfect hello-world. About a 20-minute investment of your time. And because that local model exposes a REST API, you can wire it into the things you already use, like Ansible or your own Python scripts. Empower your staff with free, private, local models, start with read-only exercises, and build from there.

Compare that to the old way. Reading your IPAM out of an old Excel spreadsheet and populating your source of truth used to be a week-long project and hundreds of lines of brittle, bespoke, throwaway code.

With MCP, you plug it in like a USB key and talk to those external systems in natural language. You’re only writing that glue code once, if ever.

What MCP Actually Is (Say It Backwards)

I came up on protocols. RIP, OSPF, BGP, HTTP, SMTP, NTP. Where would any of us be without HTTP or SMTP? So the fact that MCP is a protocol is exactly why it clicks for network engineers. And if you say it backwards, it makes even more sense: it’s a protocol that provides context to a large language model. That’s its entire job.

Under the hood it’s JSON-RPC 2.0, and it’s client-server. If I’m the MCP server and you’re the client, I advertise my tools to you: “Here are the ten tools I have available.” They plug into you, and when you need one, the model reasons about which tool to call to reach the goal. That’s the whole game.

This is the part that matters: MCP is how you let a model safely talk to the network.

It is the opposite of copying and pasting configs into a public chatbot, which is the wrong way to go about all of this. MCP gives the model visibility and a controlled way to act. It bridges the gap the right way.

From “Hello World” to Production: This Is Where Guardrails Matter

So you’ve done the local exercise. You’re talking to your network in natural language. Now you hit the real wall, the one I hit too: it’s one thing to run a read-only experiment on your laptop, and another thing entirely to run an agent against production.

The network is still the bottleneck because the network touches everything. We have to be careful with change management, change controls, and approvals. That’s not bureaucracy for its own sake. When the average cost of downtime is somewhere around $12,000 per minute, an agent that acts without controls isn’t a productivity gain, it’s a liability.

This is exactly why we built Itential FlowAI, and why we announced its general availability at Cisco Live. FlowAI is how you take that same natural-language, agent-driven approach and actually run it in production, with the controls a network demands. FlowAgents come with two non-negotiables:

1. Guardrails and role-based access control, so you decide what the agent can touch and what is off-limits.
2. A mission log that records every piece of the agent’s reasoning and every tool it called, so you have a full, auditable trail.

That mission log is the part I keep coming back to. It lets you examine what the agent actually did against the instructions you gave it. That’s the difference between hoping an agent behaved and being able to prove it did.

And here’s the part that connects straight back to your hello-world: we published Itential Builder Skills on the Anthropic marketplace. You install it into Claude Code, and through natural language you build these FlowAgents with no code. You type your prompt, the agent does its thing, and within 10 or 15 minutes you have a real, executable agent inside the platform. Same natural-language on-ramp you started with. Production-grade governance underneath it.

That’s the whole arc. Start read-only and local to learn. Graduate to governed and in production when you’re ready to let an agent act. You don’t have to choose between moving fast and not breaking prod.

This Isn’t Just About Speed

Yes, we all want to go faster. But the real prize is bigger than speed. With AI, MCP, and governed agents, infrastructure finally gets a seat at the table on business requirements and business problems. We get to deliver faster, with higher quality, with fewer mistakes, and remediate faster when there is an outage.

And for the network engineer who has always wanted to learn automation but never had the cycles, the budget, or the time? Programming used to be a multi-year investment while you were also chasing your NP, your IE, and keeping up with EVPN and SRv6. Natural language and MCP have abstracted away a huge amount of that pain.

I’m not saying don’t learn to code. I’m saying you can leap to the front of the line while you do.

Your Future Is in Your Hands

AI is here and it’s staying, but it is not here for your job. It’s here to make you a better network engineer. It’s another tool in the toolkit, and the sooner you start using it, the sooner you unlock the things you never had time for. You’re not going to lose skills. You’re going to gain them, and go down rabbit holes you never would have found.

No one is going to do this for you. But if you do it yourself, if you start building Git repos, start building projects, start standing up MCP servers, the switch in your head flips from zero to one fast. And once it does, you’re empowered by this technology, not replaced by it.

Come do it with people. The VibeOps Forum is an open, judgment-free community to share your projects, your repos, and the latest news, because community is how we actually get there. Reach out to me on LinkedIn or X. Then take your first step. Just try it. Your future is in your hands.

Watch the Full Conversation

This whole post came out of a conversation I had at Cisco Live with Dakota from The Bearded IT Dad, live in the VibeOps Lounge. We went deeper on all of it: the fear, the first steps, what MCP really is, and where FlowAI picks up when you’re ready to run agents in production. If this resonated, watch the full thing and hear it in the room where the energy was.

Watch the full interview on-demand below👇