COBRA-Skills cuts agent skill tuning cost 55-58%

COBRA-Skills cuts the cost of optimizing AI agent skills 55-58% versus SkillOpt. It also topped average performance among compared methods, per a paper on COBRA-Skills.

Most methods for tuning agent skills lean on a stronger teacher model to produce and grade candidates. COBRA-Skills doesn't need one. It worked when the target model wrote and refined its own skills, so a team without a frontier-model budget could still run the loop.

The method treats skill optimization as a budgeted, sequential search over a candidate pool that keeps changing. A bandit algorithm decides which candidates get evaluation budget next. It targets skills that look promising or informative, rather than spreading evaluation evenly across the whole pool. An evolution step then builds new candidates from evidence in what's already been tried, using 50 unique optimization examples per benchmark across six agent benchmarks and three target models. The approach held up when the paper swapped out the underlying harness.