A preprint reports that a staged AI agent led multi-turn tool tests across model sizes, with its strongest gains on state-heavy tasks but higher latency.