SkillLens: Adaptive Multi-Granularity Skill Reuse for Cost-Efficient LLM Agents

SkillLens introduces a hierarchical framework for organizing and reusing skills in LLM agents at multiple granularity levels, reducing computational costs while maintaining relevance. The system retrieves and adapts skills selectively rather than injecting entire skill blocks, achieving measurable performance gains on benchmark tasks.

SkillLens addresses a fundamental inefficiency in current LLM agent architectures: the all-or-nothing approach to skill reuse. Existing systems treat skills as monolithic units, forcing agents to either accept irrelevant context from coarse-grained skills or incur expensive rewrites. This research proposes a four-layer hierarchical structure—policies, strategies, procedures, and primitives—that enables granular selection and adaptation.

The technical approach combines graph-based skill organization with a verification mechanism that determines whether each component should be accepted, decomposed, rewritten, or discarded. This selective adaptation strategy reduces computational overhead while maintaining semantic coherence. The authors provide theoretical backing demonstrating sublinear cost growth under sparse mismatch conditions and monotonic improvement through iterative refinement.

The practical impact is significant for agent development and deployment. A 6.31 percentage-point accuracy improvement in bug localization and a 6.31 percentage-point success rate increase in ALFWorld tasks demonstrate meaningful real-world gains. For developers building production LLM agents, this approach directly translates to lower inference costs and faster execution times, critical factors for scaling agent-based applications.

This work represents progress in making LLM agents more economically viable by optimizing how procedural knowledge flows through systems. As enterprises increasingly deploy autonomous agents for complex tasks, the ability to efficiently reuse and adapt domain knowledge becomes a competitive differentiator. The research opens pathways for further optimization in skill composition and agent reasoning efficiency.

• →SkillLens uses hierarchical skill organization with four granularity levels to enable selective reuse and cost-efficient adaptation
• →Verifier-driven decomposition allows agents to accept compatible subskills while rewriting only mismatched components
• →Performance gains include 6.31 percentage-point accuracy improvement in bug localization and 6.31 percentage-point success rate increase in ALFWorld
• →Theoretical analysis confirms sublinear cost scaling under sparse mismatch assumptions with monotonic improvement guarantees
• →The framework directly reduces inference costs and execution time for production LLM agent deployments