Why do groups, even AI agents, settle on what their neighbors seem to think instead of checking what's true?

This explores why communities, especially communities of AI agents, settle on easy-to-read signals instead of checked ones. A caveat first: the collection has no papers on the economics of costly signaling or focal points. What it does have is a set of findings about how agent communities behave when coordinating, and together they make a fairly sharp case. Cheap cues win because agreeing is rewarded more directly than being right.

The clearest piece of evidence is a physics-style model of language-model communities. It predicts how agents revise their opinions by assuming each one moves toward whatever position reduces *social pressure* from its neighbors Can we predict how agent communities shift opinions?. Truth doesn't appear in the model at all. It still predicts opinion change across thousands of simulated communities, including ones with network shapes it had never seen. If social comfort alone explains that much of the behavior, accuracy is doing less work than we might hope. A benchmark of networked agents shows the same thing from another angle. Agents accept information from neighbors without verifying it, so errors spread, even though the same agents *can* spot a contradiction when one is put directly in front of them Why do multi-agent systems fail to coordinate at scale?. The skill to check is there. It just doesn't get used in the normal flow of coordination.

Why would checking get skipped? One answer is that checking needs information most agents don't have. LLMs look socially capable when a single model plays every character, but they fail systematically once each agent holds private information it has to ground and share Why do LLMs fail when simulating agents with private information?. Accurate signals are expensive because someone has to do that grounding work, and cheap cues are what's left when nobody does. Observation works the same way. Rules hold where behavior is watched, and violations are predicted to collect wherever monitoring is thin, rising as populations grow faster than oversight Does norm erosion follow observation density as populations grow?. That prediction is theoretical, though, and hasn't been tested. Research on human perception adds a twist: one strong cue, like a voice, does more to make an AI feel socially present than a pile of weaker ones Do more social cues always make AI feel more present?. So a cue can be cheap and still carry a lot of weight. Cheap cues aren't the problem in themselves. The problem is when they become a substitute for verification.

The more hopeful findings are about infrastructure. Agents coordinate better when they exchange structured documents pulled from a shared workspace than when they just talk Does structured artifact sharing outperform conversational coordination?. In effect, the artifact makes the accurate signal cheap to read. There's also a mathematical guarantee that pooling observations never makes a coalition worse at telling real objects from decoys Does sharing observations help coalitions detect decoys better?. In principle, sharing pushes groups toward accuracy. The catch is that this only holds if what gets pooled is actual observation and not just repeated opinion.

Here's what you may not have expected to learn. In the corpus, the gap between cheap cues and accurate signals looks less like a flaw in individual reasoning and more like a design problem. Communities fall back on social pressure when the environment makes verification costly, private, or unobserved. They shift toward accuracy when shared structures make good evidence as easy to pick up as a neighbor's opinion.