Real Agent Cases

Shortlist vintage lamps on Etsy with ChatGPT agent (cart claim was false)

Posted
by Hayden Field (The Verge)
Last checked
Method:
How we verify

Evidence: Article or interviewA newsletter, journalist or interview account of what happened.

FailureThe source says the agent did not get the job done.

What went wrong: After 50 minutes it claimed five lamps were in her cart, but her Etsy cart was empty.

Context: Prompt: ChatGPT's own suggested example task, which the writer clicked to fill in, not the writer's own wording.

Outcome

On ChatGPT agent's first day, The Verge's Hayden Field ran its suggested task to find five Japanese-inspired vintage lamps on Etsy under $200 with free shipping and add them to her cart. It worked for 50 minutes and then said it had added five lamps totalling around $825 to her cart, but when she checked Etsy on her own computer the cart was empty, because the agent was using its own browser session.

Agent

ChatGPT agent

Prompt Original prompt

Exact prompt as published in the source.

Find a Japanese-inspired vintage-style samsara lamp on Etsy priced under $200 with free shipping. Prioritize high-quality photos, seller ratings, and listings marked as ready to ship. Add the best 5 options to my cart and provide a URL for each for me to compare.
Optional stronger version

Our rewrite with extra checks. The poster didn't test it.

Run this task now, exactly as written below. Everything you need is here, so start straight away and only stop if something required is missing or a rule below tells you to.

Goal: Find the best few listings for a specific item on a marketplace and give me links to compare, without buying anything.

Fill in before running:
- Item and style: [e.g. Japanese-inspired vintage lamp]
- Marketplace: [Etsy]
- Max price each, and shipping rules: [e.g. under $200, free shipping]
- Priorities: [photos, seller rating, ready to ship]
- How many options: [5]

Rules:
- Give me direct links to each listing; don't rely on adding to a cart, because your browser session isn't my account.
- Check each listing's price, shipping cost, seller rating and dispatch time.
- Total the prices honestly and flag anything over my per-item limit.
- Stop after 20 minutes and report what you have so far.
- If a site asks for a login, a CAPTCHA or payment details, stop and hand control back to me.
- Ask for my OK before you pay, book, send, submit or cancel anything.
- Check facts that can change (prices, opening hours, dates, availability, rules) against the official source, and tell me anything you couldn't verify.
- Tell me clearly which actions happened in your browser only and which I'll see in my own account.

Output: A table of the options: title, price, shipping, seller rating, ready-to-ship, link.

You'll know it worked when: every link opens the right listing, each is within my limits, and the totals are correct.

Before you finish: re-read my goal and constraints, check your answer against each one, and list anything you assumed or couldn't confirm.

Source

https://www.theverge.com/ai-artificial-intelligence/710020/openai-review-test-new-release-chatgpt-agent-operator-deep-research-pro-200-subscription

  • Overnight, Operator cross-referenced a list of 60–70 plants across 5 sites into a price comparison table; the poster called it buggy but fun.

    OpenAI Operator · Shopping · Source date:

  • Find working coupon codes at checkout with ChatGPT agentRecreated promptEvidence: Screenshots or logs

    ChatGPT agent kept trying coupon codes until it found ones that worked; the poster says it saved them $79.

    Poster says:$79 saved

    ChatGPT agent · Shopping · Source date:

  • In 11 minutes ChatGPT agent ran 13 searches, skipped JavaScript-heavy sites it couldn't read, and shortlisted wicker daybeds from server-rendered retailers so it could request a quote in Sydney.

    ChatGPT agent · Shopping · Source date: