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

Abstract

Large language models (LLMs) are increasingly used to generate scientific hypotheses, and a more capable model is widely assumed to drive output quality. We tested this with a biology-focused multi-agent system coupling an LLM to five Model Context Protocol servers exposing 66 external tools. Upgrading the Synthesizer and Critic from Claude Sonnet 4.6 to Opus 4.8 initially appeared to improve quality, count, and reliability. Under a fully matched harness (identical prompts, schema, hypothesis count, and queries), the differences vanished: scores were indistinguishable on an iPSC-reprogramming task and favoured Sonnet 4.6 on a CHO cell-line-engineering task (18.4 vs 17.4; Welch t = 2.89, p < 0.05). The dominant, controllable driver of quality is therefore the prompting harness, not the model tier. We then introduce the ABCD-Stack (contract-gated prompting, diversity-first generation, adversarial refinement, and cross-family evaluation), which builds these properties in as architecture and shifts outputs toward explicitly falsifiable, causally framed hypotheses beyond the reach of a single-domain investigator. The hypotheses are falsifiable by construction but not yet experimentally validated.