Hi all, work done during my internship with Openhands (in collaboration with @wide-doctor-37830, Greg Durrett, Isil Dillig and @limited-student-10747). We worked on refining existing agent skills using traces already collected by the agent. This saves cost 1.4x-42x over other SoTA techniques while matching or beating performance on various tasks such as spreadsheet formula authoring, math reasoning, and PR review. Paper: alphaxiv.org/abs/2610.skillrefiner-offline-skill-refinement X: x.com/AnirudhKhatry/status/2106071467259543893 LinkedIn:linkedin.com/posts/anirudh-khatry_if-youve-deployed-an-agent-youre-already…?rcm=…&…

Awesome! Great work - this looks super valuable!

👀 Super cool