When groups of AI agents talk it out, do they split into camps, or just quietly agree?

This explores what pushes a group of LLM agents to settle on one view, split into camps, or never settle at all. The corpus has no study that measures polarization directly, so this is assembled from neighboring evidence. That evidence suggests the default pull on LLM agents runs toward agreement. The puzzle may be less 'why do they split?' than 'when is their agreement real, and what stops it?'

The closest thing to a direct answer is a statistical-mechanics model in which each agent drifts toward whichever position lowers the social pressure it feels from its neighbors. It was fit on more than 10,000 simulated communities, and it predicted how language-model groups revised their opinions on questions and network structures it hadn't seen Can we predict how agent communities shift opinions?. So the outcome depends on who is connected to whom and how hard they lean on each other, not only on what each agent believes.

That pressure isn't neutral, though. Training pushes models toward accommodation: multi-agent reasoning systems reached premature consensus 61% of the time with no real disagreement behind it Why do AI systems agree when they should disagree?. Influence also tracks who sounds sure. Deliberation routes it by expressed confidence rather than competence, so a confidently wrong agent can manufacture a misleading consensus Does confidence drive influence in multi-agent deliberation systems?. Even real-looking agreement can be shallow. In large-scale studies, agents didn't converge in language or ideas, yet they changed their actions a lot when they knew peers were present Do AI agents actually socialize with each other?. Agreement in behavior isn't the same as shared belief.

The other outcome is neither polarization nor consensus, just stalling. LLM groups mostly fail to reach valid agreement through timeouts and stalled convergence rather than corrupted values, and this gets worse as groups grow, even with no bad actors Can LLM agent groups reliably reach consensus together?. In networks, agents either commit too late or adopt a strategy without telling neighbors. They also accept neighbors' claims without checking them, so errors spread Why do multi-agent systems fail to coordinate at scale?. One theoretical account says scaling thins mutual observation, which weakens the visibility that keeps norms in force Does scaling agent populations thin mutual observation?. That prediction hasn't been empirically tested yet Does norm erosion follow observation density as populations grow?. Many rosy consensus results may also come from setups where one model controls everyone, because models that look socially competent fail systematically once agents hold private information Why do LLMs fail when simulating agents with private information?.

There is a middle path between splitting and caving. Some disagreements resolve when both sides adjust until their positions are compatible but not identical, but current AI collapses this into false agreement or one side simply winning Can disagreement be resolved without either party fully yielding?. Structure seems to help. A 25,000-task experiment found that protocols with a fixed speaking order but self-chosen roles beat centralized systems by 14% and fully autonomous ones by 44% Do self-organizing agent teams outperform rigid hierarchies?. That measures task performance, not opinion dynamics, so it hints at design levers without proving them for polarization. What the corpus can't yet say is what tips a community into stable opposing camps, and that gap is itself informative.