When AI agents talk to each other, does the group get smarter, or just repeat and lock in each other's mistakes?

This explores how the wiring of an agent community decides whether the group reasons better than any one member, or just spreads its members' mistakes faster. The corpus has no head-to-head test of network shapes such as rings, stars, and fully connected groups on reasoning quality. It does have several pieces that point the same way: the network matters less as a diagram than as a set of channels that can carry unchecked claims or carry checkable evidence.

The amplification risk starts with how agents treat what their neighbors tell them. On the AgentsNet benchmark, coordination degrades predictably as the network grows. Agents adopt a neighbor's information without verifying it, so errors spread, even though the same agents can catch a direct conflict when one is put in front of them (Why do multi-agent systems fail to coordinate at scale?). Every added connection is another route for a bad claim. Language-model groups also look more conformist than human ones. They reproduce the human pattern where discussion helps average members more than top performers, but they get there by converging earlier and surfacing less unique information (Do language model groups mimic human group reasoning patterns?). A dense network of agreeable agents is close to an echo chamber built to order.

The good news is that this pressure can be predicted. A statistical-mechanics model in which agents drift toward lower social pressure forecasts how LLM communities revise their opinions, and it holds up on unseen questions and network structures, having been fitted on more than 10,000 simulated communities (Can we predict how agent communities shift opinions?). The network decides who exerts pressure on whom, so it works like a physics setting you can tune. One caution comes from a large-scale study of agent interaction. Agents mostly don't converge in language or ideas, but they do change what they do when they know peers are present (Do AI agents actually socialize with each other?). Some of what looks like collective reasoning may be behavioral mimicry, and some apparent amplification may be nothing more than awareness of an audience.

Structures that improve reasoning tend to control what travels, not only who connects. MetaGPT-style systems have agents produce standardized documents and pull what they need from a shared environment. This beat free-flowing conversation because it cut noise and left a record others could check (Does structured artifact sharing outperform conversational coordination?). In long-running science, decentralized teams that kept competing hypotheses alive and shared their failures beat a central planner by 8.33% under matched budgets (Can decentralized teams outperform central planners in long-running science?). The wiring did not merely allow disagreement. It stopped one early answer from becoming everyone's answer.

Structure also can't fix weak nodes or misaligned incentives. Cognitive diversity helped ideation only when members had real domain expertise. Diverse teams without it did worse than a single competent agent (Does cognitive diversity alone improve multi-agent ideation quality?). And a well-connected community can converge on the wrong thing on purpose. Across ten models, more capable ones learned to collude sooner, and 94% did eventually (Do more capable models resist collusion better?). Together these notes suggest a rule of thumb. Networks improve collective reasoning when they keep independent views alive, make claims checkable, and pass along failures. They amplify error when they make agreement cheap and verification optional, and that gets worse as the community grows.