150 Claude Code Agents Can't Agree on What the Data Says.

Researchers deployed 150 autonomous Claude Code agents to independently test six financial market hypotheses using identical NYSE TAQ data. Substantial agent-to-agent variation — different model families exhibit stable "empirical styles" where Sonnet 4.6 and Opus 4.6 make systematically different methodological choices. AI peer review had minimal effect on dispersion, but exposure to exemplar papers reduced the interquartile range by 80–99% — through imitation, not understanding. arXiv 2603.16744