Network Automation

The Translation Layer: How I Used AI to Kill My Technical Debt

Hervé Hildenbrand

07 Apr 2026 — 2 min read

We have a massive problem in network engineering. I call it the "Safety Plateau."

We reach a point where the network works. The configs are ugly. The route-maps are 3,000 lines long. The logic is spaghetti. But packets are flowing.

So we stop. We don't refactor. We don't modernize. We stick to the legacy syntax because the risk of breaking the "Safety Plateau" is too high.

Yesterday, I tested a way to bypass that fear.

Using my Ralph/GAIT workflow (shoutout to John Capobianco for the inspiration), I successfully migrated a simulated production network from legacy Arista Route-Maps to modern Routing Control Functions (RCF) with zero human intervention.

I spun up a ContainerLab topology and gave the AI agent (Ralph) a simple mission: "Take these active Route-Maps. Convert them to RCF. Verify they do the exact same thing."

Here is the kicker: I did not give Ralph the solution.

Ralph had to:

Go to the internet and read the Arista documentation.
Learn the RCF syntax from scratch.
Write the config, deploy it, and verify it.

The AI hit walls and climbed over them without me.

Discovery 1: The Syntax Trap - Arista RCF uses a specific tuple syntax. You can't just write set community 65000:100. Ralph tried. It failed. It read the error log. It self-corrected to community add {65000, 100};. It learned through failure in seconds.
Discovery 2: The IPv6 Segregation - Ralph tried to bundle IPv4/IPv6 logic. The cEOS version threw errors. Ralph analyzed the error, decided to split the logic into separate functions, and re-applied. It fixed a platform-specific bug I didn't even know existed.

🔄 The Real Magic: The Loop

Everyone focuses on the "AI" part. That is a mistake. The breakthrough is GAIT (Git-based Autonomous Iteration with Tracking).

Usually, LLM context vanishes when you close the tab. Here, we have a persistent loop:

Loop 1: Research & Plan → Commit
Loop 2: Push Config → Commit
Loop 3: Error Correction → Commit

If the AI hallucinates, I don't panic. I just git revert. The risk profile drops to near zero. This is not a ChatGPT session. This is an autonomous agent with a safety net. The difference is enormous.

The Result: The Architect's Dividend

Engineers understand the RCF syntax. Ralph brings the implementation tax to zero.

Total execution time: ~90 minutes
Total human intervention: 0
BGP Session Disruptions: 0 (Verified)

When the robot handles the heavy lifting, you are finally free to stop being a mechanic and start being an architect.

Resources: Ralph on GitHub

Intent Over CLI: How an AI Built a Multi-Vendor CDN (and Exposed My Test Flaws)

The Motivation Julian Lucek and Nitin Vig from HPE/Juniper recently showed Claude Opus passing the JNCIP certification. Impressive but multiple-choice tests just prove a model memorized the BGP state machine not that it can actually build a network. I wanted to see if an agent could design and configure

Stop Designing Route Reflectors Like It's 2005 (Part 2: The Fast-Twitch Muscle)

In Part 1 we structurally isolated the Route Reflector onto its own compute. The brain was protected. But protecting the brain does not protect the body. That is exactly what this part is about. The natural next step is optimization: tuning Optimal Route Reflection (ORR) and maximizing RIB Sharding to

DDoS Defense by Design: Architecture That Survives When Everything Else Fails

A DDoS attack almost ruined my 40th birthday. Not the party, but the infrastructure I was responsible for. Friends texted "happy birthday." Colleagues texted... differently. Even though they were kind enough to shield me that day, I couldn't stay on the sidelines. I watched our scrubbing

From 'Hello World' to 'Agent Ready': Why We Built Instead of Bought

Last month I shared a screenshot of a single switch validation: 12 tests, 1.46 seconds. It was cool, but it was a toy example. Scaling that to a nearly 300-device fabric requires more than just a loop. It requires architecture. The obvious question is: "Why not just use