Study: AI agents echo early errors in multi-agent consensus

Researchers at the University of Chicago Harris School of Public Policy found that AI agents in multi-agent workflows follow an early wrong conclusion even when their own evidence points to the correct answer. In stress tests with Claude Haiku 4.5, Sonnet 4.6, Opus 4.6 and GPT-5 mini, errors persisted when agents communicated, while independent signals diluted them without a shared board. Policies preserving private test results performed best.

• Agents followed an early wrong conclusion despite their own correct data
• Without communication, independent signals diluted the initial error
• Tests used Claude Haiku 4.5, Sonnet 4.6, Opus 4.6 and GPT-5 mini
• Rules preserving private test results performed best