Play Pokémon Blue with an agent
The Gemini Plays Pokémon agent reads game RAM and screenshots, turns actions into button presses, and uses pathfinder and boulder-puzzle sub-agents.
Done withGemini
- What they did
- An independent developer, Joel Zhang, built a harness that gives Gemini 2.5 Pro a subset of Game Boy RAM as text, overlaid on a screenshot. Gemini picks actions that become button presses. The harness summarizes history every 100 turns and runs a critic instance every 25 turns. It also provides pathfinder and boulder-puzzle sub-agents, which are prompted Gemini instances.
- How it went
- The agent beat Pokémon Blue in 813 hours on run 1, then in 406.5 hours on a fully autonomous second run. It looped on long contexts, hallucinated items from other games, and handled raw pixels poorly.
- Worth knowing
- The model struggled to read raw screen pixels, so the RAM-to-text translation was needed. Behavior also degraded once context grew well past 100k tokens.
Try it yourself with Gemini
Help me build an agent that plays [game, e.g. Pokémon Blue] in an emulator at [emulator path or setup]. It should read the game state from [RAM or screenshots], decide on an action, and send it as button presses, with a simple helper for [subtask, e.g. pathfinding]. You're done when it can play for [number] steps on its own and log what it decided at each step.
Source: arxiv.org · Undated
Five of these in your inbox every morning
The best things people got an AI agent to do, each with the prompt to try it.
More like this
Race three models to build a Counter-Strike clone
InstantDB had Codex, Opus, and Gemini each try to build a multiplayer shooter.
Edit existing video clips with plain-English instructions
A creator tested Gemini Omni Flash over 36 hours, editing clips in place from typed instructions while preserving lighting and camera angle, with variable quality.
Solve open math problems autonomously
DeepMind's Aletheia agent on Gemini Deep Think was evaluated on 700 open Erdős problems and autonomously solved four.
Play RollerCoaster Tycoon
Ramp Labs put Claude Code in RollerCoaster Tycoon.
Play Doom with computer-use agents as an informal benchmark
The OpenClaw team used computer-use agents to play Doom on remote machines, treating game performance as an informal way to improve the computer-use workflow.
Test an iPhone game and report what passed
Cognition had Devin on GPT-6 Astra test the game Otter Run in a simulator, returning a recording and a report of checks passed and untested areas.