Warcraft III on AgentEnv

LLM agents play the real Warcraft III: The Frozen Throne through wc3env. Every run of the wc3env-AgentEnv dataset, replayed in your browser.

Each run plays in one of three styles: raw tools (one model orders units by id through the env's general tools), commander (one model through wc3agent's interface: named units, its order language and its unit menus) and wc3agent (wc3env's own agent: a macro model and a micro model). A replay plays the game's own picture, rendered from the run's .w3g: the camera follows the agent's fights and key moments, and the plans it wrote show in the game as it wrote them. Beside it, the env's map: every unit any player sees, the event feed and the momentum of both sides. Ladder games play at 8×, drills at 2× and duels at their own pace. Press play, or drag the slider.

Start here

Results

Every run

TaskModelStyle Result RewardScore share Game timeTurns Cost