What happens when models critique each other?

Ask one model to propose and another to challenge it, and something interesting happens: the thread gets smarter than either answer alone.

In our early experiments, the pattern that works best is explicit roles — one model plans, another critiques, a third synthesizes. Each sees the full conversation, so critiques compound instead of repeating.

Single-model chats hide disagreement. Multi-model threads surface it, and the user gets to watch the reasoning happen instead of trusting one confident voice.

This is the core bet behind Relay: intelligence isn't just in the models, it's in the conversation between them.