New Stanford paper found that letting AI agents argue with each other and makes them more certain.

The team ran over 10,000 small communities of language-model agents.

Letting agents discuss a math problem moves the group toward the right answer.

Each had 32 agents with different personas, linked by friendly or unfriendly ties.

They traded messages for 8 rounds about math problems with a right answer and political statements without one.

On math, the talking helped.

Wrong majorities flipped to the correct answer more often than correct ones flipped away.

On politics, the same process pushed groups rightward in 3 of the 4 models.

If your system uses agent debate, track which way the group drifted, not only whether the score went up.

– arxiv. org/abs/2608.16578

Title: "Physics of Agents: Statistical Mechanics Predicts Collective Behavior of AI Agents"