KubeCon + CloudNativeCon Europe 2025

LLMs can generate React apps, poems, and even music. But can they rise to the ultimate challenge: writing reliable Kubernetes controllers? The Config Connector team say "yes!" We are successfully using AI to write production controllers for a thousand google cloud resources.

Our path was to first break the problem into LLM-friendly steps (such as generating KRM types, the mocks and the reconciler). For each step, we invoke custom fine-tuned LLMs in a novel way with custom “jigs”. We add testing to create an “interlock” that mitigates hallucinations.

This journey changed our whole codebase philosophy: from optimizing for lines of code, we now prioritize the ability to safely and easily author and merge focused changes (at the expense of having lots of code). Although AI motivated this trade-off, it also aids development as OSS.

We believe our approach is generally applicable; join us to learn lessons that will apply as your project embraces the AI-assisted future.