A solo maintainer's account of building KubeStellar Console (a Kubernetes multi-cluster dashboard) using AI coding agents, and the hard lessons learned about making AI-assisted development reliable. The initial euphoria of fast output gave way to cascading failures, leading to a structured 'AI Codebase Maturity Model' with five stages: Assisted, Instructed, Measured, Adaptive, and Self-Sustaining. Key insights include externalizing coding preferences in instruction files, treating tests as the trust layer for autonomous workflows (91% coverage across 32 nightly suites), only automating after measurement is in place, and prompting agents with 'why' questions rather than 'fix' commands. The result: 81% PR acceptance over 82 days, with bugs resolved in ~30 minutes and features landing as PRs in ~1 hour — all without a better model, just better surrounding infrastructure.
