SkillLens Cuts LLM Agent Costs With Hierarchical Skill Reuse
arXiv.org published research on SkillLens, a hierarchical framework that organizes agent skills into four layers: policies, strategies, procedures, and primitives. Instead of injecting entire skill blocks, the system retrieves mixed-granularity components and rewrites only the parts that do not fit the current task. On MuLocbench, it delivers a 6.31 percentage-point Acc@1 gain for bug localization. On ALFWorld, agent success rate moves from 45.00% to 51.31%.
Existing skill libraries treat reuse as all-or-nothing. Inject a full skill block and you risk irrelevant context corrupting the agent. Rewrite the whole thing and you pay for tokens you did not need. SkillLens splits the difference: a verifier decides whether each visited skill unit should be accepted, decomposed, rewritten, or skipped entirely. The system also refines its own routing decisions over time.
The cost argument is the one to watch. Researchers provide theoretical backing that mixed-granularity adaptation incurs sublinear cost under sparse mismatch conditions. For any operator running agents at scale, that is the number that matters. Skill reuse is not new. Surgical, verified, partial skill reuse is.