Agent Skills Actually Hurt Performance? WebDev Benchmark Study Reveals

dair_ai · x · 2026-08-26

This paper tests 31 WebDev skills via WebDev-Skills-Bench across 50 projects. Surprisingly, injecting skills reduces Pass@2 by 1.3%-4.2% and raises token costs by 72%-394%. It identifies two failure modes: length distraction (long irrelevant prompts cause loss) and content misleading (skill content itself degrades scores). Findings also show weak skill transfer across models and concentrated losses on easy tasks.

Original post →

More from coding & agent

coding & agent channel →