A self-evolving framework to discover and refine agent skills. Most agent skills I see today are hand-crafted or poorly designed by an agent.
Multi-agent systems for building skills look promising.
This paper introduces EvoSkill, a self-evolving framework that automatically discovers and refines agent skills through iterative failure analysis.
EvoSkill analyzes execution failures, proposes new skills or edits to existing ones, and materializes them into structured, reusable skill folders.
Three collaborating agents drive the entire process.
An Executor that runs tasks, a Proposer that diagnoses failures, and a Skill-Builder that creates concrete skill folders.
A Pareto frontier governs selection, retaining only skills that improve held-out validation performance while keeping the underlying model frozen.
On OfficeQA, EvoSkill improves Claude Code with Opus 4.5 from 60.6% to 67.9% exact-match accuracy. On SealQA, it yields a 12.1% gain. Skills evolved on SealQA transfer zero-shot to BrowseComp, improving accuracy by 5.3% without modification.
I will continue to track this line of research closely. I think it's really important.
Paper: arxiv.org/abs/2603.02766
Learn to build effective AI agents in our academy: academy.dair.ai