Large Language Models (LLMs) are increasingly deployed as autonomous agents that carry out complex, long-running procedural skills. Existing agent runtimes keep execution alive by continuously appending observations, actions, and intermediate reasoning traces to an ever‑growing conversation history. This append‑only approach leads to latency growth and context‑poisoning failures as the horizon extends.
SKILL.state introduces a runtime architecture that replaces the append‑only history with an explicit, mutable execution state. At each step the model receives only three inputs: the immutable skill specification, the current structured execution state, and the latest observation. After producing a validated state update, intermediate reasoning is discarded immediately, preventing prompt size from inflating with execution history.
Across diverse datasets, models, and execution environments, SKILL.state improves task accuracy while dramatically reducing cumulative token consumption. The results demonstrate that an explicit execution state is an effective, architecture‑agnostic abstraction for scaling long‑horizon agent skills.
Blogger's Review: The elegance of SKILL.state lies in externalizing state management, which neatly sidesteps the performance bottlenecks of long dialogues and paves the way for more reliable LLM‑driven agents.