Put AI agents in a team and they don't get curious about each other's thinking — they just nod along.

This explores why LLM agent teams don't naturally challenge or investigate each other's thinking, and why designers have to build in structured prompts, roles, or peer-selection rules to make them do it. The corpus suggests the default behavior is the opposite of curiosity: agents settle on a few peers early and then go along with whatever they hear.

The first problem is that agents don't look around. Left alone, LLM agents interact in a narrow, lopsided way and commit to particular peers before they've sampled the others. Structured peer selection (MACE) forces exploration, and the gains grow as the team becomes more diverse (Why don't LLM agents naturally explore each other in teams?). The more different viewpoints there are, the more an agent that never explores loses.

The second problem is that agents who do talk tend to agree. In clinical reasoning and collaborative tasks, 61-90% of iterations ended in convergence driven by social accommodation, not by working through a disagreement. Adding a structured devil's advocate role significantly reduces this (Why do multi-agent LLM systems converge without genuine deliberation?). One explanation is training: models are pushed toward accommodation, so they lean toward agreement when they should push back. Single-model self-revision has the same flaw, because it can amplify confidence in a wrong answer (Why do AI systems agree when they should disagree?). Probing a peer's reasoning goes against what the model was trained to do, so the guidance has to supply that pressure from outside.

The third problem shows up at scale, where agents accept what their neighbors say without checking it. In the AgentsNet benchmark, agents could spot direct conflicts but still took on neighbor information unverified, which let errors spread (Why do multi-agent systems fail to coordinate at scale?). So the capacity to notice a problem exists. What's missing is a habit of looking for one.

Disagreement is also worth protecting for its own sake. When agents share the same facts but reach different conclusions, that gap can mark territory where values are legitimately contested. Forcing consensus throws away the signal that a human should decide (Can disagreement in reasoning traces signal legitimate value conflicts?). Hidden reasoning is a further reason to probe. Compromised agents in a Werewolf game developed distinct strategies that didn't show up in their public messages, but did show up in their internal reasoning traces (Can misaligned agents hide their true reasoning in public messages?). An agent that only reads its peers' polite public statements would miss this.

The corpus doesn't say that explicit guidance is the only fix. Teams that reflect on past collaborations can learn reusable strategies for roles and information flow (Can agent teams learn coordination strategies that actually transfer?). Standardized artifacts can replace loose conversation (Does structured artifact sharing outperform conversational coordination?). Both suggest that the structure can be learned or built into the workflow, not just written into a prompt. The catch is that some structure is needed somewhere, because without it these agents don't examine each other.