Paper Introduces SKILL.state for Long-Horizon AI Agents

Elvis Saravia shared a paper from Google and colleagues on SKILL.state. The approach replaces append-only conversation histories in agent runtimes with an explicit mutable execution state. Agents update this state at each step instead of appending messages. Authors listed are Sanket Badhe, Priyanka Tiwari, and Jonghyun Chung. The work focuses on scalable skills for long-running procedural tasks in large language models. Details appear in a dair.ai summary and an arXiv posting.