Researchers have introduced COBRA-Skills, a framework that reframes the optimization of reusable skills for large language model (LLM) agents as a budgeted sequential optimization problem over a dynamically evolving candidate space. The work is published as a paper on Hugging Face.
Existing skill optimization methods typically depend on costly execution-based evaluation and large amounts of task data. COBRA-Skills instead couples contextual-bandit-guided prioritization with evidence-grounded skill evolution, selectively allocating evaluations to candidates that appear promising or informative while continually refining the skill population using execution feedback.
The framework maintains a population of skills and relies on four components: a neural predictor to estimate skill utility, LinearUCB to balance exploration and exploitation, target-agent execution feedback to update the bandit, and scheduled evolution operators — regeneration, rollout mutation, and crossover — to evolve the population.
Across six heterogeneous agent benchmarks and three target models, covering search QA, spreadsheets, document understanding, mathematical reasoning, social reasoning, and embodied tasks, COBRA-Skills reportedly achieves the strongest average performance among compared methods. It reduces optimization cost by 55–58% relative to SkillOpt while using only 50 unique optimization examples per benchmark.
Additional analyses indicate the approach remains robust to changes in the agent harness, performing consistently under Codex and Claude Code, and works effectively even when the target model itself generates and refines skills. Notably, the cost savings do not mainly stem from fewer target-agent executions; a large share comes from avoiding repeated LLM-based trajectory analysis and skill rewriting.