This browser won't start 3D graphics (WebGL) right now. Try again in a minute.
You never control an agent. You build the world they learn in, press Train, and two teams of little hiders and seekers teach each other, live, until the compute runs out.
The learning is real: each team is a small neural net, trained by PPO in self-play in your browser's spare cores, the method OpenAI's hide-and-seek agents learned with (Baker et al., 2019). Nothing they do is scripted, so no two runs go quite the same way.