Google Paper Cuts AI Agent Token Usage by 94% in Long Sessions

anselm · x · 2026-09-01

Google released 'SKILL.state,' a runtime architecture replacing append-only conversation history with an explicit, mutable execution state. By providing only the skill specification, current state, and latest observation at each step, it discards intermediate reasoning to prevent prompt growth. Results show improved task accuracy and significantly reduced cumulative token consumption across diverse environments, solving latency and context-poisoning issues for long-horizon agents.

Related event: Google's SKILL.state Slashes Agent Session Tokens by 94%(2 posts)→

Original post →

More from coding & agent

coding & agent channel →