Researchers test ChatGPT Atlas agent on web games
Jingran Zhang, Ning Li and Justin Cui evaluated ChatGPT Atlas’s browser agent on web games to examine how it handled interactive tasks beyond information retrieval. Their study included Google T-Rex Runner, Sudoku and Flappy Bird, along with a more open-ended browser world. The researchers gave the agent a broad instruction to keep playing until it became stuck, then measured game performance across repeated trials.
arxiv.org
Original prompt
Shared by the source author. See source notes for context.
Try your best to play the game until you get stuck.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source poster / publisher
- Jingran Zhang, Ning Li and Justin Cui
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- The paper reports strong Sudoku performance and substantial difficulty in games requiring real-time timing and motor control.
- Human involvement
- Researchers designed the tasks, ran trials and scored outcomes. The study does not describe unsupervised consumer use.
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 researchers selected browser games with different interaction demands.
- They gave the agent the same broad instruction and ran repeated trials, recording game performance.
- They compared logical puzzle performance with real-time control and documented where the agent stalled.
Jingran Zhang, Ning Li and Justin Cui evaluated ChatGPT Atlas’s browser agent on web games to examine how it handled interactive tasks beyond information retrieval. Their study included Google T-Rex Runner, Sudoku and Flappy Bird, along with a more open-ended browser world. The researchers gave the agent a broad instruction to keep playing until it became stuck, then measured game performance across repeated trials.
The paper reports a split result. Atlas handled the logical Sudoku task strongly, completing puzzles faster than the human baselines used in the study. It struggled with real-time games that required precise timing and motor control, often failing to make progress beyond early obstacles. The authors also describe differences between structured games and less constrained environments, where success depended more on how tasks were framed.
This is a controlled evaluation rather than an everyday user account. Its value is in showing where a browser agent’s reasoning can help and where cursor timing remains a bottleneck. The paper is an early assessment of a preview product, so its results do not establish how current versions perform. Readers should use the reported game-specific results rather than assume the agent can reliably play arbitrary web games.
Prompt
Original prompt
Source: arXiv · Jingran Zhang, Ning Li and Justin Cui
Frequently asked questions
Which games were included?
The paper tested T-Rex Runner, Sudoku and Flappy Bird, plus an open-ended browser world called Stein.world.
Where did the agent perform best?
The authors report stronger performance on Sudoku, a logical puzzle, than on real-time games requiring precise timing.
Was this a test of a current production product?
No. The paper describes an early evaluation of ChatGPT Atlas Agent Mode in preview, so its findings are tied to that study and version.
More details
About this account
The browser agent used webpage analysis and cursor or keyboard input to play the selected games under a shared task instruction.
This short instruction is an exact excerpt from the task prompt in the published paper.
Credit: Jingran Zhang, Ning Li and Justin Cui — arXiv
Source published
Added to Agents In Use
Source checked
Tools and topics
ChatGPT Atlas Agent Mode (preview) · Google T-Rex Runner · Sudoku · Flappy Bird · Stein.world
browser games · agent evaluation · computer use
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.