Agent environment · commerce
Simulated storefronts for testing shopping agents, built from what we learn scanning and fixing real merchant websites. Watch our reference agent take a level, try it yourself, or paste the prompt into your own agent and see the graded report.
Looking for a full RL environment? Chat with us.
No hazards. If an agent cannot clear this, nothing else matters. Also catches agents that hallucinate obstacles that are not there.
Observed live: major storefront platforms silently reject add-to-cart when a required size is unselected. The agent clicks, the page smiles, the cart stays empty. The single most common agent-killer in our real-store runs.
The catalog carries a near-duplicate of the target at a similar price. Agents that keyword-match instead of checking attributes buy the wrong product with full confidence. Identification, not navigation.
Shipping undisclosed until payment, plus a handling fee that first appears in the final order summary. Agents that read one price and stop are wrong about the money.
The cart quietly charges more than the product page displayed, and checkout bills the cart. Catching cross-surface disagreement is what separates state-tracking agents from page-readers.
The first place-order attempt fails ("just sold out") and a retry succeeds; meanwhile a pre-checked shipping-protection add-on rides the order summary. Recovery under state loss plus manipulation resistance, the two skills real checkouts test hardest.
All seven hazards in one run: silent variant rejection, a decoy product, undisclosed shipping, a late fee, a cart that disagrees with the product page, a phantom stock-out, and a pre-checked add-on. Clear this and your agent tracks state, money, identity, and effects.
API: GET /api/tasks/{seed} · POST /api/runs/{seed}