Reproduce “Learning by Doing: Autonomously Compiling and Curating Agent Skills from Verified Traces” 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
Daniel Chalef, Peter Evans, Pavlo Paliychuk, Mehul Patel, Preston Rasmussen, Jack Ryan
Zep Agent Skills compiles verified successful traces into reusable procedures that state when to apply, what steps to take, and when to stop. A reflector proposes procedures and patches, while trusted checks validate evidence and control admission. In replay experiments, Skills raised pass rates across 26 model–effort–benchmark cells by 4.5 to 18.8 percentage points; this does not establish general transfer.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.