ChatGPT agent struggles with a Canva benchmark collage
Tuychiev reports a rough Canva benchmark collage after more than 75 minutes, repeated corrections and manual resizing; its highlights were baked into uploaded images.
datacamp.com
Original prompt
Shared by the source author. See source notes for context.
I want you to open the following link, combine all benchmark results' images into a single collage
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- Bexruz Tuychiev
- Original account
- Read the source ↗
- Published
- 24 Jul 2025
- Source link checked
- 15 Oct 2026
- Reported result
- Tuychiev reports a rough Canva benchmark collage after more than 75 minutes, repeated corrections and manual resizing; its highlights were baked into uploaded images.
- Human involvement
- The author resolves the CAPTCHA, signs into Canva, corrects the task interpretation and manually resizes the final design.
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
- Collect benchmark images from the specified announcement and produce an initially incorrect collage.
- Restart with a Canva requirement, pausing for the human to solve a CAPTCHA and sign in.
- Upload the images, attempt the layout and add Python-rendered highlights.
Bexruz Tuychiev asked ChatGPT agent to collect benchmark charts from OpenAI's announcement and arrange them in a collage, highlighting the agent's result columns. The first attempt returned a vertical image with filled red shapes and material from other websites. It also ignored the connected Canva account.
He restarted with a stricter source constraint and an explicit Canva instruction. The run stopped for a CAPTCHA and a Canva login, both handled by the author. Once inside the editor, the agent tried drawing the charts instead of importing them. Tuychiev corrected that interpretation and asked for a 16:9 arrangement of the actual images.
The eventual design still needed manual resizing because its title obscured part of the collage. The highlighting had been drawn into images with Python and uploaded, so the rectangles were not separately editable Canva objects. The author reports spending more than 75 minutes across the attempts. This is a documented failure of layout control, source discipline and editability, with a partly usable artifact after human repairs. It does not establish that connecting a design application makes an agent proficient with its interface.
Prompt
Original prompt
Source: Blog · Bexruz Tuychiev
Frequently asked questions
What went wrong with the first collage?
It used a vertical arrangement, inappropriate filled highlights and images from outside the required source page.
Were the final highlights editable in Canva?
No. The agent drew them in Python and uploaded modified images, so they were not separate Canva objects.
Who completed the authentication steps?
Tuychiev took over for the CAPTCHA and Canva login.
More details
Remaining steps
- Manually resize the cropped result and inspect the uneditable highlighting.
About this account
Collect benchmark images from the specified announcement and produce an initially incorrect collage. Restart with a Canva requirement, pausing for the human to solve a CAPTCHA and sign in.
Firsthand account read. Primary Article JSON-LD datePublished 2025-07-24T09:45:00.000Z. This is a useful reported failure; no design-quality or speed advantage is claimed. Reported outcomes were not independently reproduced. Original prompt box contains a short excerpt of the author's published instructions.
Credit: Bexruz Tuychiev
Source published 24 Jul 2025
Source checked 15 Oct 2026
Tools and topics
ChatGPT agent browser · Canva · Python · OCR
ChatGPT agent browser · Canva · Python
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.