Building Agent Harnesses for Scientific Curation from Multimodal Sources

Researchers introduce Beaver, an AI agent harness designed to extract structured information from scientific papers containing multimodal evidence (text, tables, figures). The system achieves 81.0 on the Gold-Referenced Attribute Score, outperforming frontier agents by 23 points, demonstrating that harness design—not just underlying models—is critical for complex information extraction tasks.

Beaver addresses a fundamental challenge in scientific AI: extracting nuanced information scattered across multiple document formats. Traditional language models struggle with scientific curation because the required data lives in heterogeneous sources and demands cross-modal reasoning rather than simple text extraction. This work demonstrates that specialized agent architectures can substantially improve performance on knowledge-intensive tasks.

The research builds on growing recognition that frontier AI models require thoughtful scaffolding to perform well on complex workflows. Beaver's multi-component approach—combining multimodal tooling, task staging, and artifact-based iteration—represents a design philosophy gaining traction in AI development. The iterative evaluate-diagnose-revise loop creates transparency through persistent artifacts, enabling developers to identify and fix stage-specific failures systematically.

For the AI and scientific communities, this has meaningful implications. Scientific literature curation is a foundational bottleneck limiting knowledge discovery. Improving automated curation could accelerate research velocity across disciplines. The 23-point performance gap over baseline agents suggests substantial room for improvement through intelligent harness design, not just scaling models. Ablation studies confirm that task scaffolding and multimodal evidence tooling each contribute meaningfully, validating the multi-layered approach.

Looking forward, the broader trend toward specialized agent harnesses will likely define competitive advantages in enterprise AI. Organizations deploying AI for knowledge work will increasingly compete on harness design rather than model choice. This research also hints at emerging best practices: provenance tracking, staged workflows, and artifact-grounded iteration appear critical for auditable, high-accuracy systems. Future work may extend these patterns to other multimodal knowledge tasks beyond scientific curation.

• →Beaver agent harness achieves 81.0 GRAS score, 23 points above frontier agents, proving harness design drives performance on complex tasks
• →Multimodal evidence tooling and task scaffolding are essential components for extracting information from text, tables, and figures simultaneously
• →Artifact-grounded iteration enables transparent, auditable workflows where failures localize to specific stages for targeted fixes
• →Cross-modal reasoning and data normalization show the largest performance gains, indicating complexity concentrates where modalities intersect
• →Agent harness design emerges as a central competitive lever for knowledge-intensive AI applications beyond just model selection