AI coding agents can generate code from natural language, but most workflows assume a human who can read and review that code. What happens when the human cannot? This experience report documents a 45-day project in which I—a PharmD/PhD statistician who cannot read Python code—built an operational MCP server (~190k LOC including tests) using AI agents with zero human-written code. I could not inspect the implementation, so code review was impossible. After an initial failure where specification-driven development produced modules that "looked right but did not run," the project pivoted to an ADR-driven control plane governing a fast-changing data plane (code, tests). The control plane—decision logs (ADRs), policies (rules), procedures (commands), and quality gates—stabilized inter-module contracts and coordination. I describe this control plane's minimal components, how failures drove its evolution, and practical patterns for collaborating with AI agents without inspecting code.
Katsuya Shibuki (Tue,) studied this question.