Google's SKILL.state Tames Long-Horizon Agents

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.