Can you build and ship a real site almost entirely with AI agents?

Most "AI builds it" demos collapse on contact with production. The thesis: with the right human-in-the-loop seams, agents can ship real software.

The pain point: everyone shows agents writing toy apps, then quietly hand-edits the real thing before the demo. We wanted the honest version. Build and ship this actual site with agents doing the volume and a human holding the seams.

Where does it work, where does it fall apart, and what is the orchestration pattern that survives production? This idea is the spine of the build-log research, and eventually the open-source tooling that comes out of it.