Three researchers proposed a new way to run AI agents on long, multi-step tasks. Their design, SKILL.state, replaces the ever-growing chat history with a structured record of progress. At each step the model sees only the task specification, that record, and the newest observation. Intermediate reasoning is thrown away once it has produced a validated update. The authors say this stops the prompt from growing as work continues. They report better task accuracy and much lower total token use across several datasets, models and environments. They argue the approach does not depend on any particular model or system design. The paper was accepted at EMNLP and posted to arXiv, with a revised version on 28 August 2026.