Google Paper: Replace Agent Chat History With Explicit State, Cut Tokens 16x

rohanpaul_ai · x · 2026-08-29

A new Google paper, SKILL.state, argues long-horizon agents need far less conversation history than we give them: swapping the ever-growing transcript for a small explicit execution state cuts tokens and often improves accuracy.

At each step the model receives only skill instructions, a structured state, and the latest observation; reasoning traces are discarded after producing a validated state update, keeping prompt size roughly constant. On a 100-step warehouse task, Gemini-3-Flash with SKILL.state scored 0.94 using 65,408 tokens vs 0.91 and 1,062,387 tokens for a LangGraph-style baseline — a 16.2x difference. Boundary: everything needed later must fit in the state schema.

Related event: Google Paper: Explicit State Cuts Agent Tokens 16x(2 posts)→

Original post →

More from coding & agent

coding & agent channel →