Agent skills are reusable, actionable natural-language artifacts that guide an agent to perform a task effectively under a given harness. While millions of such skills have been shared publicly, most existing skill optimization methods largely ignore this accumulated knowledge and instead rely on costly agent rollouts to iteratively refine a skill for a target task.
To address this gap, a new paper proposes Retrieval-Augmented Skill Optimization (RASO), a framework that uses an external skill corpus as prior knowledge throughout the optimization process. RASO retrieves relevant knowledge from existing skills and adapts it to the target task and harness through a mechanism the authors call Cross-Harness Adaptation, which accounts for mismatches in both domain and harness.
RASO operates in two complementary stages. Retrieval-Augmented Skill Initialization (RASI) constructs a knowledge-grounded initial skill without requiring any agent rollouts, while Retrieval-Augmented Skill Update (RASU) iteratively refines the skill by retrieving external knowledge guided by execution feedback.
Across four agent benchmarks and two language models, the authors report that RASO consistently outperforms baselines that lack retrieval-augmented skill initialization and updating. The results also indicate that directly reusing retrieved skills can be ineffective—or even harmful—when their tools and assumptions do not match the target environment, making Cross-Harness Adaptation a key component.
The broader takeaway, according to the authors, is that publicly shared skills can serve as transferable prior knowledge for agent improvement—not by copying them verbatim, but by retrieving and adapting the right procedures.