SKILL.state paper reported: 94% token reduction and accuracy gain on 100-step Gemini 3 Flash benchmark
Following·since 29 Aug 2026·Day 2·2 sources
Google SKILL.state cuts agent token use 94% while improving accuracy
The gist
Long-horizon agents currently accumulate large context windows that grow with task length; SKILL.state keeps prompt size roughly constant as tasks grow longer. The benchmark shows lower token use alongside higher accuracy, though the method is bounded by what the state schema can represent.
The full picture
Researchers from Google and Purdue published SKILL.state, an approach for long-horizon AI agents that replaces the growing conversation transcript with a compact structured state. At each step, the model receives only the skill instructions, the current structured state, and the latest observation; reasoning traces are discarded after each step. On a 100-step warehouse benchmark using Gemini 3 Flash, SKILL.state used 65,000 tokens versus 1.06 million for a LangGraph-style baseline, a reduction of roughly 94%, while accuracy improved from 0.91 to 0.94. The approach requires that all information needed in later steps can be captured in the state schema.
How it developed
Sources
Rohan Paul Twitter
The daily email
Want this in your inbox?
I send a short email each morning with the stories that moved. If you would rather just read here, that works too.
Subscribe free