A room holds one agent session rather than one per person. When somebody asks for work, everyone present watches the reply arrive token by token, and anyone can interrupt it mid thought and send it in a different direction.
That changes what a review is. Instead of one person receiving output privately and relaying a summary, the whole room sees the same reply, the same hesitation, and the same mistake at the same moment. Disagreement happens while the work is still cheap to change.
Calls are peer to peer over WebRTC, so audio and video never touch our servers. The call limitation is stated below, because peer to peer has a real failure mode and it would be dishonest to print only the privacy half.