AI Beyond Autocomplete: Using LLMs To Create 1000 Kubernetes Controllers
Talkat KubeCon + CloudNativeCon Europe 2025 · on Friday, 2025-04-04· at 11:45· Level 0 | ICC Capital Hall | Room I·
summary
Claims building 1000 controllers for Google Cloud resources by running custom fine-tuned LLMs at build time to generate KRM types, mocks, OpenAPI schemas and reconcilers across about 12–15 steps. Pipeline uses reusable “jigs and interlocks” to validate outputs (compile, HTTP logs, 404 checks) and runs in parallel — e.g., first pass succeeds on ~600/1000 resources — producing commit-ready controllers with tests and human review, and the approach is reported as working in production.
AI-written summary · about
chapters (10)
- 0:00AI and Config Connector intro
- 3:14LLMs for code generation
- 7:46Non-deterministic LLM outputs
- 12:01LLM strategies and Bitter Lesson
- 15:39Simple vs complex LLM tasks
- 19:56Parallel task validation
- 23:08Scaling controller generation
- 26:18Trust in LLM-generated code
- 28:46Automation and takeaways
- 31:38Conclusion and discussion
related talks
31mCluster Operations as a Service, Introducing LLM Backed Controllers.
37mBeyond Kubebuilder - Generating Entire Kubernetes Controller Implementations
30mLLMs on Autopilot: Running AI Agents on Kubernetes with Open Source Tools
42mLLMs on Autopilot: Running AI Agents on Kubernetes With Open Source Tools
14mAI Deployment: Mastering LLMs with KFServing in Kubernetes
43mForget Kubectl and Talk to Your Clusters: Using LLMs to Simplify Kubernetes Cluster...
30mTaming Dependency Chaos for LLM in K8s
14mReasoning Operators: Bringing LLM Logic Into Kubernetes Control Loops