A new framework, Retrieval-Augmented Skill Optimization (RASO), optimizes agent skills by leveraging an external skill corpus and adapting it to the target task and harness. RASO outperforms baselines in extensive experiments across four agent benchmarks and two models.