Building Agent Harnesses for Scientific Curation from Multimodal Sources

The article introduces Beaver, a new agent harness designed for structured scientific curation from multimodal sources, which effectively extracts information from diverse formats such as text, tables, and figures while maintaining provenance. Beaver integrates a frontier agent with multimodal evidence tooling and task scaffolding, achieving a Gold-Referenced Attribute Score (GRAS) of 81.0, surpassing previous frontier agents by over 23 points. This development highlights the importance of harness design in enhancing agent performance for scientific workflows, particularly in tasks requiring cross-modal reasoning.