Design the system. Direct the agents. Govern the output.
I am not a classically trained software engineer and I don't pretend to be. I design architectures, contracts and guardrails, then direct AI coding agents to implement them — under a supervisor with one writer per repo, deterministic evals, and a human approval gate on every write. That workflow produced a quarter-million lines of tested code in six months. The skill that matters now is knowing what to build, how to verify it, and when to say no.
Specify precisely
Contracts, data shapes, failure modes and acceptance checks written before any code. Agents get a spec, not a vibe.
Run agents under governance
Work-unit graphs, one writer lease per repo, prompt-hash grants, an AI reviewer, then a human gate. Nothing merges on confidence alone.
Prove it with evidence
Deterministic evals, golden journeys, architecture claims that must cite proof, and metrics that render NOT_AVAILABLE rather than lie.