De NETCONF al Vibe Coding — y el dilema del que nadie habla

We lived through the network automation era.

NETCONF. RESTCONF. YANG models. Cisco DNA Center. Python scripts built on ncclient and netmiko. Ansible playbooks. Staging labs. Idempotent configs. It was powerful — but it demanded deep expertise. You had to know the protocols, the data models, the libraries, AND the network itself before a single line of automation touched production.

Now AI is changing the game again.

With tools like Claude Code, GitHub Copilot, deepseek TUI and Codex, you can describe a complex network automation workflow in plain English and get working code in minutes. Need a Python script that connects to your multi-vendor MPLS fabric via RESTCONF, pulls every VRF routing table, and flags missing default routes? Describe it. Done.

This is vibe coding applied to network infrastructure.

But here's the dilemma nobody is being honest about:

With traditional network automation, you OWN the code. Every YANG path, every policy-map, every edge case — you reasoned through it, tested it in staging environment, and understood exactly why it works. The slowness of the process enforced discipline. By the time that script touched production, it had a provenance.

With vibe coding, the knowledge gap between idea and runnable code collapses. But the gap that actually matters — between runnable code and correct, safe, production-grade code — stays exactly as wide as it always was.

AI can generate a route-map reference that IOS-XE accepts silently but never applies. It can produce a YANG path valid on IOS-XE 17.9 that doesn't exist on 17.3. And because the code appears so quickly, there's a dangerous temptation to skip the staging environment entirely.

The quality of your output is also only as good as your prompt. If you can't articulate that you need a hierarchical QoS policy with a parent shaper and MQC child classes applied per-subinterface — you won't get a correct result from any AI model.

So what's the right mental model?

Treat AI-generated network automation code the way a senior engineer treats a junior's pull request — valuable, potentially correct, but requiring rigorous review before it touches production.

The prompting is the briefing. The review is your professional judgment. The staging test is the non-negotiable gate.

AI doesn't eliminate the need for deep network knowledge. It raises the ceiling for what a knowledgeable engineer can build — and raises the floor for how much damage a less knowledgeable one can cause.

The tool doesn't decide which outcome you get. You do.