When every assistant agrees
Panels of AI agents are increasingly used to simulate stakeholders, debate hard questions and make forecasts. They agree far too easily. When agents share a model, share their evidence, or can see each other's answers, their agreement says little about whether they are right.
Three ways agents end up agreeing
They are the same model. A city could replace public workshops with fifty simulated residents, each given a different persona, and get a report saying most of them back the plan. In a study of 120 such deliberations, agents assigned different value perspectives still converged on the same option. Giving each perspective a different underlying model cut that concentration sharply: in a child-welfare scenario, from 71% picking the same first choice to 46%.1 Personas written into a prompt are not the same as different people.
They copy each other. In debates between identical models, agents adopt the majority answer, abandon correct reasoning when shown a peer's argument, and lose correct answers in the final vote. In one study, isolated self-correction beat group debate.2 A replication of the classic Music Lab experiment found large populations of language-model agents far more prone to herding than people, and less able to recover from a runaway cascade.3
They see the same evidence. Forecasting agents given identical evidence end up herding instead of revising their beliefs, and the group does little better than one agent.4 The same thing happens outside AI panels: outlets that rely more heavily on the same wire copy publish sooner after the leading outlets do.5
Each of these produces agreement that looks like confirmation and isn't.
What we propose
- Different models for different perspectives. Where there is no single correct answer, such as a policy consultation, model diversity preserves disagreement that one model playing many roles erases.1
- Private evidence. Give each agent some information the others don't have, so their errors are less correlated and deliberation has something to add.4
- Targeted exchange instead of broadcast. Have one agent respond to a specific claim of another (support, rebuttal or question) rather than showing everyone the running tally.6
- Less chatter. In distributed search, agents that share beliefs constantly converge early on wrong answers. Sharing only when something has changed protects exploration.7
- Report the dissent. A simulation's output should show minority positions and how strongly they were held, not only the winning option.
The objection
For questions with one right answer, diversity alone doesn't help much: the same study that found model diversity reduced convergence in policy questions notes that accuracy-oriented debate behaves differently.1 There, the useful checks are independent verification of the answer itself (see why an AI agent will not audit itself), not more voices.
What's unsolved
- No simulation of residents should replace asking residents. We don't know how to validate simulated stakeholders against real ones.
- Even with private evidence and tools, joint decisions under partial information remain hard for current models.8
- How much model diversity is enough, when most available models are trained on overlapping data.
Related: Why many small assistants beat one giant mind.
Sources
1. Ariel Sela, "Preserving Disagreement: Architectural Heterogeneity and Coherence Validation in Multi-Agent Policy Simulation" (2026), arXiv:2604.26561.
2. Blaž Bertalanič and Carolina Fortuna, "The Cost of Consensus: Isolated Self-Correction Prevails Over Unguided Homogeneous Multi-Agent Debate" (2026), arXiv:2605.00914.
3. Farnoosh Hashemi, Shiyu Ji, Joice Chen and Michael W. Macy, "LLM Multi-Agent Systems Are Vulnerable to Herding and Runaway Cascades" (2026), arXiv:2610.30769; the original experiment is Matthew J. Salganik, Peter Sheridan Dodds and Duncan J. Watts, "Experimental Study of Inequality and Unpredictability in an Artificial Cultural Market," Science 311 (2006): 854–856, doi:10.1126/science.1121066.
4. Yuante Li, Yicheng Tao, Kate Zhang et al., "Diverse Evidence, Better Forecasts: Multi-Agent Deliberation Under Information Asymmetry" (2026), arXiv:2607.01661.
5. Mohammad Yousuf and Arif Md Tareque Habib, "Structural Dependence, Imitation or Herding? A Multi-Agent Simulation of Global Intermedia Agenda Setting" (2026), arXiv:2610.20144.
6. Dezheng Bao, Sihang Du, Renhong Huang et al., "Beyond Answer Sharing: TrueMAD for Targeted Peer Engagement in Multi-Agent Debate" (2026), arXiv:2610.18240.
7. David Farr, Iain Cruickshank, Kate Starbird and Jevin West, "The Cost of Consensus: Malignant Epistemic Herding and Adaptive Gating in Distributed Multi-Agent Search" (2026), arXiv:2605.06988.
8. Chenxu Wang, Yongkun Yang, Boyuan Du et al., "LLM Agents for Deliberative Collaboration: A Study on Joint Decision Making Under Partial Observability" (2026), arXiv:2607.06157.
