Google's EfficientSkill prunes agent skills in two phases, cutting output tokens up to 80%

dair_ai · x · 2026-11-17

A Google/Purdue paper introduces EfficientSkill, a two-phase framework for token-efficient multi-turn agents: first repair failures with no length penalty until target accuracy, then prune instructions and move reusable code into workspace utilities while accuracy stays close to target. Across four models on MATH-500, ALFWorld and SpreadsheetBench, output tokens fell in 11 of 12 configurations (up to 80%) while accuracy rose on every benchmark (+14.3 pts on Gemini 3.8 Flash, +17.5 on Grok 4.1 Fast), lowering total token cost with or without full-prefix KV caching.

Original post →

More from coding & agent

coding & agent channel →