SKILL.state throws away the conversation history: only the skill spec, current state, and latest observation reach the model

arXiv·high signalXBlueskyLinkedInCopy link

arXiv 2608.26263 (Badhe, Tiwari, Chung, submitted 2026-08-26) replaces the append-only conversational history in agent runtimes with an explicit mutable execution state. At each step the model sees the immutable skill specification, the current structured state, and the newest observation only; intermediate reasoning is discarded the moment it produces a validated state update, so the prompt never grows with execution history. The paper reports higher task accuracy alongside substantially lower cumulative token consumption across datasets, models and environments, which is the opposite trade from compaction where you pay accuracy for length.