When should AI agents actually stop and think?

Large language models waste computation by planning on every problem equally. This work splits agentic reasoning into three parts: a world model that simulates future states, a learned controller that decides *when* to plan and how deeply, and a reactive executor for immediate actions. The team built SR²AM, which uses reinforcement learning to train a 30B model to match reasoning performance of models 20–30× larger while cutting reasoning tokens by 76–95%. The key insight: optimal agents don't deliberate constantly—they learn to plan only when it matters, spending extra thinking time on genuinely hard problems rather than easy ones.