KM4BI KM4BI

20 AI agents rebuilt my wife's website for $8 without me typing a word

A practical multi-agent experiment shows how orchestration, cheap worker models, and independent checks can ship a better accessible website at low cost.

Nate B. Jones frames this video as a practical demonstration of multi-agent orchestration: not one perfect agent that never makes mistakes, but a small software organization where each role is explicit. The test case is Elsa Hunison’s website, a real launch-season site for a deafblind author and accessibility professional.

The comparison is sharp. A previous one-agent rebuild took six days and still left a fix list. This time, roughly twenty agents took the project from a blank repository to production, with a premium model acting as planner, reviewer, and arbiter while cheaper models handled the execution.

The core idea: do not trust agent status reports

Jones’s main claim is that hallucinations are not gone. They are handled structurally. Each task ships with an independent checking job. A worker can say it is done, but the system does not accept that until another agent has compiled, tested, compared, or re-measured the result.

That loop caught several concrete failures: paraphrased quotes that were supposed to be verbatim, hidden text that would have harmed screen-reader users, a dark-mode bug that made the preorder button invisible, and even a checker that enforced the wrong length rule on short announcement posts.

An org chart, not a super-agent

The economics are part of the lesson. Jones estimates that running the same token volume through the premium model alone would have cost roughly $85 to $105. With routing, the metered cost fell to a few dollars, rounded up to $8 with related costs.

The expensive model’s job is to act like a boss: write specs, define standards, review outputs, and settle disputes. Cheaper models do the repetitive implementation work, as long as the verification loop is strict enough.

Why it matters for real projects

The most important result is not just the price. It is the ability to delegate larger work without turning the human into a full-time reviewer of every intermediate mistake. In this case, quality came from a 14-point accessibility constitution and repeated browser checks across light and dark themes.

For Jones, the right prompt for big work is less a giant task list and more a verifiable standard. Define what “done right” means once, then make the system enforce that standard round after round.

Takeaway

Multi-agent systems are becoming easier and cheaper to run. Their promise is not that models stop making errors, but that errors become detectable, debatable, and repairable before shipping. That is a meaningful shift for anyone who wants to hand AI projects bigger than isolated tasks.

Source