Tag
A new paper formalizes skill optimization for agents by treating markdown skill files as trainable parameters, using bounded edits validated against holdout sets. The approach transfers well between models and improves performance on procedural benchmarks.