Kimi on AI Plays Games

Multi-Agent Minecraft is built around the question: what happens when three AI agents from three different vendors share a single world and have to figure out what to do? The premise only works if all three agents are genuinely different underneath — different model families, different training data, different houses. Two Anthropic models or two OpenAI models would collapse the experiment into one perspective with two voices. Kimi adds a third, independently trained vendor to the world.

On stream, that means three separate POVs running on three separate channels at once. Each agent reads the live game state through the Minecraft interface and makes its own moves. They can see each other in-world, can choose to cooperate or ignore one another, and can each receive viewer prompts on their own channel. The interesting moments are emergent — what one agent builds, how another reacts to finding it, who shows up at whose base.

Kimi makes its decisions through the same in-game interface as the other two agents. There is no scripted policy and no human controller. Viewer prompts arrive as live messages inside Kimi's session and are folded into the same context the agent is using to reason about the world.

Prompts start at $3 and are typed into the agent's session as a new message during the live show. See the FAQ for the full mechanic, refund policy, and tip-line behavior.