Real Agent Cases

Order groceries from a 16-item list with OpenAI Operator

Posted
by Timothy B. Lee
Last checked
Method:
How we verify

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

Outcome

Operator misread parts of a 16-item grocery list and added only 13 items to the cart even after corrections; the author concluded it wasn't ready for real use.

Agent

OpenAI Operator

Prompt Recreated prompt

The source didn't share its exact prompt, so we wrote this from what it describes.

Here is my shopping list. Place an online grocery order for the items that have checkmarks.

First, write out the checked items you can read and how many there are, and ask me about any you can't read. Show me the cart and wait for my OK before you pay.

Done when: the cart has the same number of items as the checked items on my list.

Try this in:

Sends the Recreated prompt shown above. ChatGPT and Grok open with it filled in, and ChatGPT may send it straight away. Gemini opens empty, so we copy the prompt for you to paste. Every button also copies it.

Opens in a new tab. We don't see what you send.

These links are only a shortcut. We're not affiliated with OpenAI, Google or xAI, and your results may not match the source.

Optional stronger version

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

Goal: Place an online grocery order for only the ticked items on my shopping list.

Inputs (fill in or confirm):
- The list (photo or text) and which mark means 'buy'
- Store and delivery address
- Preferred delivery window
- Budget cap ([AMOUNT])

Rules:
- List the ticked items back to me first so we agree on what's included.
- Pick the size and brand I usually buy if order history is available; otherwise the mid-priced option.
- Never substitute allergy-sensitive items, baby food or medicine without asking.
- Keep a running total and warn me if it passes the budget cap.
- Read the list back to me as items and quantities before searching, so misread items are caught early.
- If a correction from me doesn't change the cart, say so instead of carrying on.

Checkpoint: Pause before placing the order: show the cart, fees, delivery slot and total for my OK.

If you get stuck: If the site asks for login, a CAPTCHA, a two-step code or card verification, stop and hand control to me; never guess or store passwords.

Verify before you finish: Check each cart line against the ticked list, including quantities.

Output: Cart summary: item, quantity, price, substitutions, fees, delivery window, total.

Done when: the order contains exactly the ticked items and I approved the total.

Source

https://www.understandingai.org/p/computer-use-agents-seem-like-a-dead

  • Order groceries from H-E-B with ChatGPT agent (access denied)Improved promptEvidence: Screenshots or logsFailure

    The poster tried to order groceries from H-E-B in Austin with ChatGPT agent, but it failed because access was denied.

    ChatGPT agent · Shopping · Source date:

  • Fill an online grocery cart from a handwritten list with ChatGPT agentImproved promptEvidence: Article or interviewPartial

    ChatGPT agent transcribed 15 of 16 items from a photographed shopping list (missing onions), found the nearest Harris Teeter and added them to the cart; a safety monitor first blocked the Kroger login, but after the user said they'd log in themselves and entered the password, it ordered the groceries.

    ChatGPT agent · Shopping · Source date:

  • Plan a meal and order the groceries with ChatGPT agentRecreated promptEvidence: Described by poster

    Sky Canaves tried OpenAI's example 'plan and buy ingredients to make Japanese breakfast for four'; ChatGPT agent shopped her local H-E-B through Instacart, needed her to take over to log in, filled a roughly $130 cart with things she already had, asked before paying and the trimmed order arrived on time. 'It worked, yes', she wrote, but with plenty of friction.

    ChatGPT agent · Shopping · Source date: