Test a browser agent on web games
As described in the source, on each game's page in the ChatGPT Atlas browser, the researchers turned on Agent Mode (Preview) and gave one zero-shot instruction: play until you get stuck. They ran 10 trials per game and did not intervene. Sudoku took an average of 2m28s with 100% accuracy, against a human baseline of 10-12 minutes.
- Agent or tool
- ChatGPT
- Topic
- Productivity
- Reported on
- Blog
- Added
- 2026-10-15
Why it matters
Small, repetitive tasks eat more of the week than anyone likes to admit. Cases like this show an agent taking a chore off someone's plate end to end, not just suggesting what to do. It is one of 196 sightings filed under Productivity on Agent Sightings (number 68 by interest), and one of 29 involving ChatGPT.
How you could try it
Here is a reusable prompt, written by us in our own words, that captures the task. Paste it into ChatGPT or any similar agent and adjust the details to your situation.
Help me test a browser agent on web games. Break it into steps, show your work, and wait for confirmation before anything irreversible, paid, or outbound.
Tip: Start small: run it once on a low-stakes example, read what the agent did, and only then let it run on real work.
Source
Read the original first-hand account on Blog: https://arxiv.org/html/2510.26298v1. Our summary is a paraphrase; details are in the source.
Built an AI agent? Feature your agent on Agent Sightings, free submission or a featured spot.