One day, one fleet, one shipped site — with receipts.

This site is its own first case study. It went from an empty Webflow project to the production deploy you are reading in one working day, built by a fleet of AI agents — image models, video models, coding agents — under one human director. This note records what that actually looked like, because the polished result hides most of it.

What ran

• The visual world (the ivory-and-brass instrument, its crew, the opened orb) was generated from owned reference images: five stills and two videos, roughly 180 generation credits end to end. Every asset carries a provenance sidecar — model, prompt, date — checked into this repository next to the file.
• The scroll section on the home page scrubs an eight-second render, re-encoded with a keyframe on every frame so the browser can seek to any scroll position without stutter. It carries three independent fallbacks: reduced motion, small screens, and a stall watchdog.
• The site shipped twice in one day: first assembled inside Webflow through its API, then — after the platform’s limits showed up in practice — ported the same evening to a static build on Vercel, where it now lives in version control like everything else we run.

What broke

• Two agents edited the same page concurrently, and the second one silently deleted the first one’s centrepiece section minutes before a publish. No history, no conflict warning — that is what a design tool without version control does under parallel agents, and it is the single strongest reason this site now lives in git.
• A media-blocking browser extension in our own test browser suppressed the scrub video without firing an error event — the request never reached the network, and every error-based fallback stayed silent. The fix that survives this failure class is a timeout watchdog, not an error handler. If your video fallback only listens for error, it has a blind spot.
• An asset-indexing run wiped a previous index because a scan flag reset state by default. Derived data, fully recoverable — but the same lesson as everywhere else: defaults decide outcomes when agents move fast.

What it means

None of this required a team. It required a system: named roles, bounded scopes, gates that check outcomes rather than effort, and receipts at every step. That system is what this lab teaches and what the rest of this site is about. The failures above are not embarrassments — they are the reason the method exists, and publishing them is the method working.

Written with AI assistance; reviewed and edited by Frank Riemer. Evidence: this repository’s commit history and the asset provenance sidecars in /assets.