Grail Hunter

Hunts a short list of specific dream watches across three marketplaces, and never bids.

A watchlist is not a shopping list. Measure intent, not interest.

Four specific watches I actually want. The AI agent searches three marketplaces for them, judges whether a listing is authentic from the photographs, and messages me when something clean appears inside budget.

How it is built

A session loop rather than a scheduled task, it runs, throttles itself, and wakes up again. Each pass: search each source, filter to real matches, open the promising listings properly, read the photographs, and score.

The authenticity pass matters more than the price pass. A dial that says the wrong depth rating, a case back with the wrong reference, a bracelet that does not belong to the model, these are visible in the listing gallery and almost nobody looks.

It is report-only by design. It has never placed a bid and the code to do so does not exist. That is a deliberate constraint: the moment an AI agent can spend money unattended, every bug becomes expensive.

What it got wrong

It treated saving a listing as wanting it. The AI agent logged every item I bookmarked as demand, and recommended accordingly. In reality bookmarking is how I keep an eye on something I have no intention of buying. Dozens of saves, zero actual bids, the AI agent had been learning from a signal that meant nothing.

It kept re-suggesting things I had rejected. Removing a bookmark is a decision. The AI agent read it as an absence rather than a no, and cheerfully surfaced the same watch again a week later. Deletion had to be made explicit and permanent.

The lesson

Most AI agent mistakes are not technical. They are a wrong theory about what a human action means. The code worked perfectly; it was the interpretation of “saved” that was broken, and no amount of testing would have caught it, only noticing that the recommendations felt wrong.