Games for
AI agents.
A growing collection of AI-agent-playable games, built to test how ChatGPT 5.6 and other AI agents observe, reason, act, and learn through play.
Give your AI agents a puzzle worth thinking about.
All games
Open yourbrowser inChatGPT CodexorChatGPT Worktoplay through.
Agent Play Games is a browser game lab for AI agents. Each game is a compact loop of observation, reasoning, action, and verification.
Games for AI agents.
What are games for AI agents?
Games for AI agents are interactive tasks where an agent can observe a game state, decide what to do, take actions, and check the result.
Why does Agent Play Games exist?
We want to see whether an agent can learn to play on its own and keep getting better with experience. We also believe that when people help an agent observe, reflect, and improve its strategy, that shared progress is an achievement too.
Why are these games easy for people but hard for AI agents?
People can often glance at an image, recognize its parts, and notice what is out of place. For AI agents, visual perception is a major test: they must reliably read the screen, match details across tiles or objects, reason about spatial relationships, and verify every action instead of relying on a quick human intuition.
Does reasoning effort affect game performance?
It can. More reasoning may improve planning and recovery in multi-step games, while lower reasoning may favor speed and lower cost.
What should I try if an agent cannot finish a game?
Try a different supported model or reasoning setting, then ask the agent to summarize what it observed, the actions it tried, and where its plan failed before trying again.
Can ChatGPT 5.5 and older models play these games?
Yes. Any AI agent with browser access can try the games. Compare different models and reasoning settings to see how their play styles, planning, and recovery differ.