Google paper details EfficientSkill for pruning agent skills
The order of optimisation — accuracy before length — is the practical takeaway, since pruning an unlearned skill stalls. Teams running agentic workflows with rising token bills can apply the same two-phase loop to their own skill libraries, cutting serving cost by a quarter to two-thirds. The +14.3 and +17.5 point accuracy gains on SpreadsheetBench for Gemini 3.8 Flash and Grok 4.1 Fast give a concrete benchmark for comparing it against other skill-optimisation approaches.