The agents look at the scene on the whiteboard in front of them.

One agent says a word which distinguishes one object in the scene from all the others.

If the other agent doesn't know the meaning of the word, the game fails.

If one agent doesn't know the meaning of a word used by another, it can't pick out the object that the other agent is talking about. But it can try to work out a property - such as a color or size - that identifies one object in the scene, and learn that as the meaning of the unknown word.