Vinted Sweep

The second-hand-marketplace twin of the auction sweep, tuned to one brand and a different currency.

Fixed-price is a different game to auction. The asking price is the price.

A sibling of the auction sweep pointed at a fixed-price marketplace in a different country and currency. Same shape, completely different economics.

Why it needed its own rules

On an auction site the listed price is a starting point; the close is what matters, and it lands well above it. On a fixed-price marketplace the asking price is the price. That single difference invalidates every threshold inherited from the auction AI agent, a lot that would be a bargain at auction is simply the market rate here.

So this AI agent buys much lower relative to market, and it narrowed to a single brand. A broad search returned a large pool that was mostly irrelevant; a narrow one returns fewer, better candidates.

What it got wrong

Brand lived in a field the AI agent never read. Matching on the listing title missed items where the brand was only in a structured attribute, so genuine in-budget pieces were logged as “not a target brand” and thrown away.

The filter let through things that are not watches. Parts, straps, boxes, clocks, all listed under a watch brand. Each needed its own vocabulary, in several languages, and the fix for one kept revealing another.

A “zero new results” run looked like a bug. It wasn’t. The catalogue genuinely turns over slower than the schedule runs. Teaching the AI agent that nothing new is a valid outcome took longer than teaching it to search.

The lesson

Cloning an AI agent to a second source is where the hidden assumptions surface. Half the logic turned out to encode facts about the first marketplace that nobody had ever written down.