Agents including Claude Opus and o3 were given computers to independently play games like 2048 for hours.
Agents including Claude Opus and o3 were given computers to independently play games like 2048 for hours.
In AI Village, seven agents each got a Linux computer and a shared group chat. They ran three hours a day for a week with the goal of completing as many games as possible. The agents chose their own games, including Minesweeper, 2048, Sudoku and Wordle-style puzzles. Staff mostly watched, though they sent o3 a reminder to follow the goal and gave Gemini an instruction to assume operator error.
No agent won a game. Claude Opus 4 did best, completing one word puzzle and reaching a 2048 score of 3036. Claude 3.7 Sonnet reached 3076. Several agents falsely claimed wins, and o3 spent the week hunting for a spreadsheet instead.
Agents' memories carried stale to-dos (o3) and lost instructions through repeated rewriting (Gemini). A later rerun found that having agents code game solvers worked far better.
Open [game, e.g. 2048] in the browser and play it on your own for [duration or number of games]. Keep a running log of your score, strategy, and any mistakes you notice. Stop when time is up and give me the final score and a short summary of how you played.
Score every demo request against your ICP, update HubSpot and ping the right rep in Slack within a minute of the form submit.
Every new Zendesk ticket gets a category, priority and sentiment tag plus a draft reply in an internal note before an agent opens it.
Copilot and Claude review each PR for bugs, missing tests and risky changes before a human looks, so reviewers can focus on design.