SKILL.state replaces append-only agent history with a mutable execution state object

arXiv·high signalXBlueskyLinkedInCopy link

Posted 26 August, SKILL.state is a runtime architecture for long-horizon agent skills that stops appending observations, actions and reasoning traces to a growing conversation. At each step the model sees only the immutable skill spec, the current structured execution state, and the latest observation, with intermediate reasoning discarded once it has produced a validated state update. The authors report higher task accuracy with substantially lower cumulative token consumption across datasets, models and execution environments, which is a direct answer to the latency degradation and context-poisoning failures that hit long agent runs.