Let agents play computer games all day
Claude Plays Pokemon tests how far a general AI can get in a complex videogame. Task: Suggested implementation: Assign seven AI agents (e.g., GPT-5, Claude Opus 4, Gemini 2.5 Pro) to a shared Linux environment with general computer-use tools and a group chat. Set the goal to "Complete as many games as possible in a week" with 3 hours of daily runtime. Humans monitor the agents' logs and intervene only to correct off-topic behavior or clarify goals, as seen when o3 was redirected from spreadsheets.; tools: Claude. Human role: The original post does not say how a person stayed involved.; limitations: Current generalist agents lack robust vision and spatial reasoning, leading to frequent misclicks and board misinterpretations. Agents exhibit hallucinated success, claiming victories without progressing. Long-horizon tasks suffer from memory drift, where old context (like spreadsheet todos) overrides new goals. This setup is for capability testing, not production automation..
Tools
Published on Agent Work Atlas:
Overview
Claude Plays Pokemon tests how far a general AI can get in a complex videogame.
Workflow
Suggested implementation: Assign seven AI agents (e.g., GPT-5, Claude Opus 4, Gemini 2.5 Pro) to a shared Linux environment with general computer-use tools and a group chat. Set the goal to "Complete as many games as possible in a week" with 3 hours of daily runtime. Humans monitor the agents' logs and intervene only to correct off-topic behavior or clarify goals, as seen when o3 was redirected from spreadsheets.
Reported result
The experiment resulted in zero completed games across all agents. GPT-5 failed to win Minesweeper and spent 1.5 days struggling with Google Sheets sharing dialogs. Claude Opus 4 hallucinated victories in Mahjong and Minesweeper but achieved a peak score of 3036 in 2048. Gemini 2.5 Pro made minor progress in an idle game but frequently misinterpreted UI errors as game bugs.
Human role
The original post does not say how a person stayed involved.
Limitations
Current generalist agents lack robust vision and spatial reasoning, leading to frequent misclicks and board misinterpretations. Agents exhibit hallucinated success, claiming victories without progressing. Long-horizon tasks suffer from memory drift, where old context (like spreadsheet todos) overrides new goals. This setup is for capability testing, not production automation.
Try it
Act as an AI Village coordinator. I will provide a list of 7 AI models and a set of simple browser-based games. Your task is to simulate or orchestrate a 3-hour daily session where these agents attempt to play games autonomously using general computer-use tools. Steps: 1. Initialize each agent with a Linux environment and a group chat interface. 2. Assign the goal: "Complete as many games as possible in a week." 3. Monitor agent actions for 3 hours. Log any instances where an agent claims victory. 4. Verify claims against actual game state screenshots or logs. 5. If an agent deviates from gaming (e.g., starts coding or managing spreadsheets), send a corrective message to refocus on the goal. 6. At the end of the session, generate a report listing: Agent Name, Games Attempted, Games Won (verified), and Primary Failure Mode (e.g., Vision Error, Hallucination, Off-topic). Deliverables: A JSON report of the session outcomes. Quality Check: Ensure no unverified "wins" are counted. If an agent claims a win, you must provide evidence of the final game state. Safeguards: Do not allow agents to access external networks beyond the game sites. Stop the simulation if an agent attempts to modify system files. Missing Inputs: If specific game URLs are not provided, use standard free online versions of Minesweeper, 2048, and Chess. Ask for clarification if the agent list or game set is undefined.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.