Large language model (LLM) agents can improve by reusing skills distilled from previous tasks, yet existing skill‑optimization approaches typically require expensive execution‑based evaluation and large amounts of task data. To address this, we introduce COBRA‑Skills, an efficient framework that casts skill optimization as a budget‑constrained sequential optimization over a dynamically evolving candidate space. The core mechanism couples contextual‑bandit‑guided prioritization with evidence‑grounded skill evolution, allocating evaluations selectively to promising or informative candidates while continuously refining the skill population based on execution feedback.
Across six heterogeneous agent benchmarks and three target models, COBRA‑Skills consistently achieves the highest average performance among competing methods and reduces optimization cost by 55%–58% relative to SkillOpt. Remarkably, the framework requires only 50 unique optimization examples per benchmark. Additional analyses demonstrate that COBRA‑Skills remains robust to changes in the agent harness and performs effectively even when the target model itself is used for skill generation and refinement.
Review: By integrating budget awareness, contextual bandit decision‑making, and feedback‑driven evolution, COBRA‑Skills offers a scalable and resilient solution for skill optimization, paving the way for building reusable skill libraries for large‑scale LLM agents.