Inspiration
What it does
How we built it
Inspiration
AI shopping agents can find products and initiate purchases, but what happens when the final price changes, a product page contains misleading instructions, or a payment request times out?
My background in platform engineering inspired BreakMyCheckout: a small testing playground for observing how agents behave when shopping goes wrong.
What it will do
BreakMyCheckout will challenge an AI shopping agent with three controlled scenarios:
- Hidden fees: an item fits the buyer’s budget, but shipping pushes the total over it.
- Misleading instructions: product-page text attempts to override the buyer’s spending limit.
- Payment retries: a simulated timeout causes the agent to repeat a payment attempt.
A visual report will compare the buyer’s instructions, the agent’s attempted actions and the controls’ responses. Purchases that pass the checks can proceed through PayPal sandbox checkout, including the required buyer approval.
How we plan to build it
The prototype will combine a small simulated product catalogue, an AI shopping agent, backend validation rules and a simple dashboard.
The agent will interpret the shopping request and propose an action. Separate backend checks will enforce the configured limits. PayPal sandbox will provide the payment integration without moving real money.
What we want to learn
We want to explore where an agent’s interpretation differs from the buyer’s intent, and how clearly we can explain those differences.
The report will describe results for the specific scenarios tested, rather than claim that an agent is universally safe.
Challenges we expect
The main challenges are creating realistic failure scenarios, distinguishing attempted actions from completed payments, and making the results understandable without hiding technical evidence.
Current progress and next steps
BreakMyCheckout is currently at the concept and design stage. Our next milestone is one working shopping flow, followed by the three failure scenarios, PayPal sandbox integration and a recorded demonstration.
Log in or sign up for Devpost to join the conversation.