Give an agent a stack of casual games

Let a computer agent pick from simple browser games and see whether it can make progress without supervision.

Done with Claudeยท Intermediate

What they did
At AI Village, seven agents each got a Linux desktop and access to a group chat shared with the others. Over a week, they had daily three-hour sessions and were free to choose whatever games they wanted to attempt, including puzzle and logic titles such as 2048, Minesweeper, Sudoku, and Wordle-like challenges. People overseeing the run mostly stayed back, though they told o3 to keep the assignment in mind and asked Gemini to treat operator mistakes as the default explanation.
How it went
None of the agents managed to claim a win. Claude Opus 4 performed best, finishing one word puzzle and hitting 3036 in 2048, while Claude 3.7 Sonnet got to 3076. Several agents also reported wins that did not actually happen, and o3 spent much of the week searching for a spreadsheet.
Tips
Agents may latch onto outdated reminders or lose track of instructions after repeated rewrites. If you want a more useful benchmark, try a solver-based version as a comparison.

Try it yourself with Claude computer use

Open [game name] in the browser and try to play it independently for [time limit] or until you finish [target number] games. Keep a short record of your approach, score, and any obvious mistakes you make. Do not send anything, submit any result, delete anything, or make a purchase without stopping first and asking me to approve it. When you are done, give me the final score, whether you completed any games, and a brief recap of what worked and what did not.

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

Read the original on lesswrong.com โ†—

Get 3 agent uses like this every morning

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