Reproduce “Scientific Agent Skills: A Library of Procedural Knowledge for Research Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Timothy Kassis, Vinayak Agarwal, Yuhuan He, Darshil Patel, Aubrey M. Brueckner
The authors present Scientific Agent Skills, a library of human-readable procedures that coding agents can load on demand. A host keeps skill descriptions resident, then reveals instructions and referenced files when relevant. The release contains 163 skills across 16 areas, with descriptions using 14,246 tokens, or 7.1% of 200,000. The paper measures documentation and lexical overlap, not scientific accuracy.
Key insights
- Explicitly authored procedural knowledge, formalized into 'skills,' can directly address pervasive scientific errors in agent-generated analyses by guiding agents on appropriate methods and caveats.
- A tiered knowledge loading strategy significantly defers context window costs, allowing agents to access a large and specialized knowledge base (83.7% unread until activation) without exceeding token limits for multi-skill tasks.
- Lexical similarity analysis of skill descriptions, combined with explicit cross-references, aids agents in distinguishing and routing to relevant procedural knowledge, although some highly similar pairs still lack direct pointers.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.