Use a browser agent to test simple games

A browser agent was asked to play several games with minimal guidance, showing strong results on Sudoku and weak ones on fast-action titles.

Done with ChatGPT· Beginner

What they did
The researchers opened each game in ChatGPT Atlas, enabled Agent Mode (Preview), and gave a single instruction to keep going until it could no longer proceed. They repeated this process 10 times for each game, without taking over, and watched the agent infer the controls on its own. In Stein.world, they then tried again with extra instructions that named the keys and the first goal.
How it went
Sudoku was solved in an average of 2m28s with perfect accuracy, compared with a human benchmark of roughly 10–12 minutes. T-Rex Runner averaged 45.5, and Flappy Bird ended at zero in every attempt. Stein.world only got past the first room after the added guidance, and the agent still spent a long time thinking before each action.
Tips
This was a small study, so the results should be treated as early evidence rather than a final verdict. Games that depend on precise controls or step-by-step directions may need more explicit setup.

Try it yourself with Operator

Open [game URL] and play the game until you can’t move forward. Learn the controls from the page if needed, then tell me how long it took, what score you reached, and where you got blocked. Before you create an account, enter payment details, submit anything, or make a deletion that can’t be undone, stop and ask for my approval.

Fill in the [brackets]. Agents can make mistakes, so review anything before it is sent, paid or deleted.

Get 3 agent uses like this every morning

With the day's top AI news. Free, and one click to unsubscribe.