A shopping agent that finds, compares and buys is the whole course in one transaction: tools, protocols, memory, evaluation, and, above all, the human gate. Every other week you could fix a bad run by running it again. This week the run moves money, and “mostly works” stops being an acceptable answer.
shop-agent, and install langsmith. The steps and the add-on download are at the top of 03 · Trace it.place_order() is a return, a chargeback, a phone call, and sometimes a loss. So: retries need idempotency (a retried order must not become a second order); the loop needs a hard spend cap enforced in the harness, not requested in the prompt; the trace stops being a debugging convenience and becomes the evidence that answers “who authorized this?”; and the human gate moves from a nice-to-must the thing the system is designed around. Reliability that was “good enough for a draft” is now the difference between a product and a liability.A small task (“restock the espresso pods for the break room, under $60, here by Friday”) touches nearly everything you have built since Week 1. Click each stage to see which week it came from.
Each stage of a purchase maps onto a layer you have already studied. The new part this week is what happens at the stage where money moves, and what must be true before you let the agent get there.
Everything the merchant needs to trust the sale is bundled into one human sitting at one screen. Fraud controls are built around that assumption.
The click is gone, so consent must be represented explicitly, as something a merchant can check. That is precisely what the new payment protocols on page 01 are trying to standardize.
An office-supplies agent with a budget. Press Play to watch a purchase flow through an intent mandate, a cart, a cap check and a payment credential, then watch the same agent misread a catalog listing. The question the week keeps asking: at each arrow, what can go wrong here, and who pays when it does?
Try this: in the second scenario the agent is never “wrong” in its own trace. Point to the single control that caught the error, and say what would have happened without it.
Not “what does the model cost” but what does one completed task cost: tokens per step, times steps, times attempts, plus the human minutes spent reviewing and fixing. Until you can quote that number, you cannot compare the agent to the person doing the job today.
Trust here is a set of enforced limits. A spend cap, an allow-list of merchants, an approval step on anything irreversible, and a kill switch. Each one is a line of code somebody can point at.
When a purchase is disputed, someone must be able to reconstruct it: what the user authorized, what the agent saw, what it chose, and what it paid. Your Week 2 evidence standard (state before → observation → available actions → selected action → result → state after) becomes a financial control.