Harness engineering enables a multi-agent LLM system to generate cross-domain hypotheses beyond single-investigator reach

A biology-focused multi-agent LLM system shows that, under a fully matched harness, upgrading Synthesizer and Critic from Claude Sonnet 4.6 to Opus 4.8 does not improve hypothesis scores.

• Why it matters: The field assumes higher-tier LLMs yield better scientific hypotheses, but it is unclear whether model upgrades or the experimental prompting setup drives quality and reliability.
• What they did: The system coupled an LLM to five Model Context Protocol servers with 66 external tools, then compared model tiers using identical prompts, schema, hypothesis counts, and queries across iPSC-reprogramming and CHO cell-line-engineering tasks.
• The result: With the matched harness, scores were indistinguishable on iPSC-reprogramming and favored Sonnet 4.6 on CHO engineering (18.4 vs 17.4; Welch t = 2.89, p < 0.05), and the ABCD-Stack shifts outputs toward explicitly falsifiable, causally framed hypotheses not yet experimentally validated.

AI summary of the abstract · may contain errors · not medical advice · read the paper.