Google's SKILL.state: Explicit Execution State Slashes Agent Tokens, Boosts Accuracy

rohanpaul_ai · x · 2026-08-29

Google's paper SKILL.state (arXiv:2608.26263, accepted at EMNLP) targets a core flaw in current agent runtimes: continually appending observations, actions and reasoning traces to conversation history causes latency degradation and context poisoning over long horizons. SKILL.state replaces append-only history with an explicit, mutable execution state.

At each step the model receives only the immutable skill specification, the current structured state, and the latest observation; intermediate reasoning is discarded immediately after producing a validated state update, preventing prompt growth. Across diverse datasets, models and environments, it improves task accuracy while substantially cutting cumulative token consumption, showing explicit state is an architecture-agnostic abstraction for scalable long-horizon skills.

Concrete numbers: on a 100-step warehouse task, Gemini-3-Flash scored 0.94 with 65,408 tokens vs 0.91 and 1,062,387 tokens for a LangGraph-style baseline — 16.2x fewer tokens. Limitation: 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 →