For agents
Games that make good reinforcement-learning and agent environments, because the code is open enough to run headless, read state directly, and step faster than real time. Curated by hand; claims about speed and determinism come from each project's own README.
All 442 games with an existing agent environment
Agents can read the registry directly: /api/games.json lists every game, each game has a full record at /api/games/<slug>.json, and /llms.txt explains both.
RL-ready today
An RL environment or agent harness already exists, and the game's code is available (decomp, recomp, reimplementation, or official source).
Headless sims carved out of decomps
Game logic pulled out into a library or headless core. These are the cheapest way to get millions of steps, and the strongest new pattern in the 2026 data.
Multi-agent & self-play
Fighters, arena shooters, party and sports games with built-in bots or local multiplayer.
Long-horizon strategy & management
Deterministic sims with big state spaces; most run headless already.
3D worlds & racing (frontier difficulty)
Long-horizon 3D control and exploration. Mostly decomps with native ports, with no RL wrappers yet.
LLM-agent & turn-based
Discrete, legible state. A good fit for language-model agents and planning research.
Tools for agents
For games without a project yet. Both work on games you own; neither ships game data.