Amazon paper: stale agent skills can drop performance 10pp and raise cost 40%

rohanpaul_ai · x · 2026-11-20

Amazon's new paper When Agent Skills Go Stale studies how agent skills behave after environment changes. Skills store past experience (file paths, field names, commands) to amortize inference cost, but silently go stale when apps or APIs change.

Findings across 4 task settings and 2 models (Claude Opus 4.7 and Haiku 4.5):

The authors propose selective maintenance: replaying checks from successful executions without extra model calls — verifying whether observed differences conflict with skill state, repairing affected content, validating updates, and rebuilding when local repair fails. vs. reusing stale state, maintenance improves accuracy by up to 67pp and cuts execution steps by 37%, with no description of the change supplied.

Related event: Amazon Paper Shows Stale Agent Skills Can Drop Performance by Over 11 Points(2 posts)→

Original post →

More from coding & agent

coding & agent channel →