This talk walks through my personal learning project that replicates a production-like Kubernetes fleet: managed end-to-end with Crossplane, GitOps powered by FluxCD, and structured around proper control-plane/workload separation and multi-tenancy patterns. This project served me as a turnkey platform where I tested ideas on a setup more sophisticated than a local kind cluster or a standalone managed cluster with default settings.

But when AI came along, initially as Claude chat, then Claude Code, then Claude agents, hooks, and eventually an agentic loop I built for this stack - I realised I had replaced myself out of the picture with an AI ecosystem I had built myself (well, with AI obviously). Like that proverbial frog that didn't notice the water getting hot, my project turned into an AI learning playground: while I was developing the agentic loop, the AI replaced Crossplane providers and performed a major version upgrade of Crossplane.

While I am still deciding whether I love this or hate it, one thing is for sure: I learned a lot about AI in the process, and I want to share these learnings in this session: context engineering, token cost, agent harnesses, and much more.