AI can now automatically extract scientific procedures from papers and code to build reusable agent skills

Researchers built a system that crawls scientific papers, code repositories, and databases to automatically extract procedural knowledge and convert it into packages that AI agents can execute. This means AI systems designed for scientific work can now learn new tasks from existing documentation instead of requiring humans to manually teach them what to do.

Scientific knowledge is scattered across papers, scripts, databases, and APIs — most of it invisible to AI agents because nobody encoded it in a form they can understand. This system closes that gap by automatically mining that knowledge and converting it into executable packages. The result is a library of 71% novel skills that existing hand-crafted libraries didn't have, and agents using these skills performed better on five of six scientific benchmarks tested. The structural change: AI agents stop being limited by what humans explicitly taught them and start learning directly from the scientific record.