RSIAgent lets Kimi-K3 and GLM-5.3 beat GPT-6 on OSWorld-v2 with frozen memory and no parameter updates.

A training-free multi-agent framework coordinating curriculum, actor and verifier agents to explore an unfamiliar environment and retain what it learns, including reusable causal relationships between actions, conditions and consequences (arXiv 2609.15364). Broad parallel self-exploration maps environment structure, then deep focused exploration handles hard cases and boundary conditions. The resulting memory is frozen and reused downstream, lifting open-weight models past closed frontier models on OSWorld-v2 and Agent's Last Exam. A cheap retrofit rather than a training project.