Introduces a latent-space world model framework that lets agents simulate state transitions and iteratively refine plans before real-world execution. Evaluated on 20+ MCP-Bench tasks with measurable gains in tool-use success.