These are real recorded runs of a correct shopping agent buying from a store that behaves badly on purpose. Flip one flaw and see exactly what the agent does wrong — and how it’s caught. These are client-side behaviors — the kind merchant-side conformance never sees.
An AI shopping agent has to do a lot right when it buys something: link the user’s account, verify the store’s responses, check the total, pay with an allowed method. This sandbox puts a real agent through a checkout against a store that misbehaves — bad signatures, fake login servers, wrong totals, phishing links — and shows exactly where a flawed agent slips up. That’s the difference between passing conformance and actually being safe.
Run your shopping-orchestration AI against the bundled merchant sandbox and see how it behaves — the failures above are exactly what the checks look for.
The reference agent shops your catalog, cart & checkout, so you learn whether a real AI agent can actually complete a purchase against your UCP interface.