On its turn the agent runs thousands of the four-phase loop you stepped through on the Chibany tree — but here each simulate is a random play-out of a whole tic-tac-toe game to a win, loss, or draw. The numbers it builds up:

Yellow number on a square = how many of its searches tried that move (it plays the most-searched one); top-left % = its estimated win rate from there. Beating it is hard — that's the point: the same loop that solved a 3-state toy plays a real game.