What they did

The authors developed COBRA-Skills, a framework for optimizing reusable instructions or procedures distilled from an agent’s previous task experience. A contextual bandit—a decision algorithm that balances testing uncertain options against exploiting promising ones—selects which skill candidates receive costly execution-based evaluations.

The framework uses the resulting execution evidence to generate and refine its candidate pool. It was evaluated on six heterogeneous agent benchmarks with three target language models, using 50 unique optimization examples per benchmark.

Key findings

• COBRA-Skills achieved the strongest average performance among the compared methods across all six benchmarks and three target models.
• It reduced optimization cost by 55–58% relative to SkillOpt.
• The method operated with only 50 unique optimization examples for each benchmark.
• Additional tests found that performance remained robust when the agent harness changed and when the target model generated and refined its own skills.

Why it matters

Reusable skills can improve agents without repeatedly solving every task from scratch, but evaluating candidate skills through full executions is expensive. The authors show that selectively allocating this evaluation budget can make skill optimization more data- and cost-efficient while accommodating a candidate pool that evolves during optimization.

Caveats

The abstract does not report absolute performance gains, benchmark-level results, variance, or the precise accounting behind optimization cost. The evidence is also limited to six benchmarks, three models, and comparisons chosen by the authors; broader testing is needed to establish how well the approach transfers to longer-running agents, different skill formats, and substantially larger candidate populations.