Google's SKILL.state Tames Long-Horizon Agents
Google researchers propose SKILL.state, a runtime that replaces growing agent transcripts with immutable skill instructions, structured execution state, and the latest observation. It reaches 0.94 accuracy at 100 steps while using 16.2× fewer tokens than a stateful baseline.
SKILL.state is a compelling runtime redesign: the agent’s operational state becomes canonical, while transient reasoning and stale history disappear after each validated transition.
- –Keeps prompt size roughly bounded as task horizons grow, reducing latency and inference cost.
- –Validated state patches filter irrelevant telemetry and enable immediate recovery from external world-state changes.
- –Budget-matched truncation and statistical compression performed far worse, suggesting semantic structure matters more than merely shrinking context.
- –Schema design becomes a critical engineering surface; weak models often overwrite, delete, or misformat state.
- –Discarded history limits provenance and debugging, while dynamically discovered schemas and multi-agent concurrency remain open challenges.
DISCOVERED
1d ago
2026-08-31
PUBLISHED
1d ago
2026-08-31
RELEVANCE
AUTHOR
Discover AI