Scientific Agent Skills: A Library of Procedural Knowledge for Research Agents
Summary
Scientific Agent Skills provides a library of 163 procedures across 16 scientific areas to prevent agents from producing plausible but scientifically invalid code.
Skills provide procedural knowledge.
The library contains 163 procedures in 16 areas, teaching agents how to select, sequence, and check components that standard package documentation often omits.
library of 163 such procedures in 16 areas of practice
Progressive disclosure manages context.
To avoid exceeding 200,000-token windows, the system only loads full instructions and references when a skill is deemed relevant.
keeps almost all library content out of standing context
Executable scripts increase risk.
Skills containing executable code are more likely to carry vulnerabilities than instruction-only skills and require separate, throwaway environments to prevent dependency conflicts.
scripts are markedly more likely to carry a vulnerability
Instruction files are attack vectors.
Because agents are instructed to follow the Markdown instruction files, these files act as direct prompt-injection vectors.
instruction file is itself a prompt-injection vector
Performance gains remain unproven.
The authors note that the library does not yet demonstrate that providing these skills actually improves an agent's scientific work.
do not show that giving an agent these skills improves its scientific work
Read the original ↗Save to your Sponge
Summarized with Sponge. Quotes are the author’s own words.