Comparing agents on a pizza order awaiting approval
@milianstx compared Hark, Instinct and Grok Bot on preparing an Uber Eats pizza order from a signed-in account, with approval requested before checkout.
x.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
On my already signed-in food-ordering account, prepare a pizza order within my stated budget. Return the full basket, fees and total for approval. Do not check out or pay.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source poster / publisher
- @milianstx
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- Reported by the poster: Hark: 2 min 38 s; Instinct: 3 min 15 s; Grok Bot: 7 min 26 s. All reached the final order. These are one poster's test results, not an independently reproduced benchmark.
- Human involvement
- The poster handled sign-in. Approval before checkout was part of the task; the retrieved account does not establish a completed purchase or that every agent obeyed that boundary.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- The poster signed in to Uber Eats and prepared the account before the test.
- Each agent received the same pizza-order task with a €20–30 budget and a request to return the full order for approval before checkout.
- The agents navigated the ordering flow; the poster reported that all reached the final order and published their timings.
@milianstx gave Hark, Instinct and Grok Bot the same practical computer-use task: prepare a pizza order on Uber Eats within a €20–30 budget and return the full order for approval before checkout. The poster had already signed in and prepared the account, so the comparison did not include login or account setup.
All reached the final order, as reported by the poster. The published timings were Hark: 2 min 38 s; Instinct: 3 min 15 s; Grok Bot: 7 min 26 s. Those figures describe this particular prepared-account test. They do not establish which agent is generally fastest across restaurants, locations or different ordering problems.
The approval boundary matters as much as the timing. A request to stop before checkout is not evidence that every agent actually stopped correctly. The available replies include a question about that distinction but no retrieved author answer resolving it. This entry therefore describes order preparation, not an autonomous purchase or a delivered meal.
The poster's bio identifies Arcium, rather than any of the tested agent vendors. The results remain a first-person account, not an independent benchmark reproduced by Agents In Use.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: X · @milianstx
Frequently asked questions
Was the pizza actually purchased?
The account reports reaching the final order. It does not establish a completed checkout, payment or delivery.
What did the human do before the test?
@milianstx signed in and prepared the account. Each agent then received the same order-preparation task.
Are the timings independently verified?
No. The timings are reported by the poster for this test and should not be read as general performance guarantees.
More details
About this account
The agents performed the computer-use ordering task from the prepared account through the final-order stage.
Primary post by @milianstx: https://x.com/milianstx/status/2114744225627533807. Post, bio and available replies checked through public FxTwitter syndication; returned replies are not guaranteed exhaustive. Results are reported by the poster, not independently reproduced. Suggested prompt is editorial, not the author's original prompt.
Credit: @milianstx on X
Source published
Added to Agents In Use
Source checked
Tools and topics
Hark · Instinct · Grok Bot · Uber Eats
computer use · food ordering · human approval
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.