From the CLI to Natural Language: A Network Engineer’s Case for VibeOps

I’ve spent my career at the command line.

Senior network engineer at an insurance company. Senior network architect for the Parliament of Canada – 50 buildings, three data centers, 500 remote sites, national importance on the line. The CLI was where the work happened. It’s where it still happens for most of us.

I recently joined Todd Kane on the Evolved Radio podcast to talk about exactly this – where network automation and AI intersect, what VibeOps actually means in practice, and why I think infrastructure engineers don’t have to be the last ones through the door this time. Here’s the expanded version of that conversation.

How We Got Here

When I first got access to ChatGPT 3.5 in November 2022, I was curious. When I got an API key and started connecting AI to my network automation work, I was obsessed.

I’d already been deep in network automation for years – self-published a book on it, co-authored a Cisco Press title on network automation frameworks. I’d seen what was possible when you stopped configuring devices by hand and started thinking programmatically. But AI was different. It wasn’t just faster execution of the same tasks. It was a different kind of interface entirely.

One night, two ideas collided in my head. I already knew how to put JSON into a vector store for RAG – retrieval augmented generation. And I knew you could use T-shark at the command line to turn a packet capture into JSON. So I asked the obvious question: could you just upload a PCAP and talk to it?

Turns out, yes. The AI scored 20 out of 20 on packet capture exams – real online exams with real questions and the answers attached. Now, with the major hyperscaler models, you just upload your PCAP to ChatGPT and start chatting. Capability that didn’t exist three years ago is now table stakes.

That’s how fast this is moving. And that’s exactly why network teams can’t afford another 10-to-15-year lag.

What VibeOps Actually Is

Vibe coding gave non-developers access to code. You apply your own knowledge, your own context, your own experience – and you use natural language to drive the outcome. The syntax is abstracted. The barrier drops.

Why does the network have to be different?

That’s the core of VibeOps. It’s not AIOps – that’s machine learning over big datasets, supervised and unsupervised learning applied at scale. Different thing. VibeOps is the idea that you can use natural language to interface with infrastructure. That instead of dropping into a CLI to check something, you just ask: How’s the health of the Wi-Fi in Vancouver? What’s the traffic doing between New York and Singapore right now?

It’s using things like MCP – Model Context Protocol – to plug the right tools into the right agents. It’s developing the skills your agent has, and using natural language to achieve goals that used to require deep technical syntax knowledge just to approach.

The thing I keep coming back to: this is more accessible than network automation ever was. You don’t need to learn Python. You don’t need to learn Ansible or Terraform. You describe what you want. You build the agent. You plug in the tools.

We were always locked out by complexity. That lock is coming off.

The Part Nobody’s Calling an HR Problem

Jensen Huang said the IT department will become the HR department for AI. People made a meme out of it. But after building and deploying agents, I keep coming back to it – because it’s genuinely true.

You wouldn’t hire a junior engineer and put them on a BGP change on day one. You’d start them with read-only access, documentation, minor tickets. You’d train them on corporate policies, change management guardrails, environment-specific constraints. You’d increase their autonomy as they earned it.

Agents work exactly the same way. Human in the loop. Human on the loop. Human in the lead. Fully autonomous. The failure mode isn’t the technology – it’s treating it like a technology problem when it’s actually an organizational one. Where do these agents fit in your org? Whose work do they augment? How do you expose them securely? Do you need supervisor agents? How high up the org chart does the hierarchy go?

That’s not an infrastructure question. That’s an architecture and HR question. And the sooner infrastructure teams get comfortable with that frame, the better they’ll use what’s actually available right now.

Start Where It’s Safe

Alert triage is one of the most underrated first use cases for AI agents in network operations. If you’ve got 10,000 tickets, point an AI at them and say: boil this down to 100. Find the patterns. Identify the repeatable issues. Correlate the noise.

You’re not touching anything. No change windows. No risk. Just signal extraction from data you already have. And from there, you start layering – suggested remediation steps, test plans, order of operations. The agent grows into more autonomy as you trust it.

For MSPs managing large client portfolios, the opportunity is even more specific. You can start building per-tenant digital representatives – agents scoped to that client’s environment, connected to your knowledge base, your tickets, your documentation. Think of it as a practical employee you’re building up and attaching tools to, one client at a time.

A few hard rules: don’t do shadow AI. This needs a proper API key, a private LLM, and real enterprise-level agreement. The YOLO DIY approach doesn’t belong near production infrastructure.

But you can absolutely start somewhere safe. Download VS Code. Copilot is built in, and there are enough free calls to get going. Get comfortable with the IDE. Then identify which MCPs map to the tools you use every day – Salesforce, Jira, Atlassian, your monitoring stack. Plug them in. That’s the on-ramp to building a real personal assistant for infrastructure operations.

The Question About Junior Engineers

I get this a lot: if agents handle Level 1 work, what does the entry-level path look like?

My honest answer: those juniors can be accelerated. They don’t have to spend 25 years getting to expert-level the way I did. The agents built by senior engineers are a mechanism for transferring knowledge – passing the torch faster and more deliberately than the informal apprenticeship model that defined previous generations of network careers.

The accounting analogy is the one I keep coming back to. Were there more or fewer accountants after Excel became widely available? More. More spreadsheets, more data, more accountants – just working at higher levels, on more meaningful problems. The job changed. Accounting didn’t disappear.

When I automated the first major thing at Parliament, nobody said thanks, see you later. They asked what we could automate next. I became more valuable. Not less. That’s what I’ve seen every time, and I expect AI to follow the same pattern – if you’re building the agents and not just consuming them.

If it’s your agent, you’re not going anywhere. If you’re only using someone else’s agent, that’s worth thinking about.

The Network Doesn’t Have to Be Last This Time

I’ve watched every wave of developer tooling leave network teams behind. It took us a decade and a half to catch up on network automation. I’ve been building in public the whole time – books, GitHub repos, the VibeOps Forum – because I think practitioners sharing their actual journeys is how we close the gap faster this time.

The interface is changing. Natural language is becoming a first-class way to interact with infrastructure. Not eventually. Now. The tools are here. The models can do it. The only question is whether network engineers get in early and shape how this develops – or spend the next decade catching up again.

I know which one I’m choosing.

Join the VibeOps Forum →

Listen to the Full Conversation

Catch the full episode on Evolved Radio with Todd Kane below – we go deep on packet capture AI, the agent autonomy ladder, local vs. cloud LLMs, and where VibeOps is headed.