How to Build Efficient Agentic Reasoning Systems by Dynamically Pruning Multiple Chain-of-Thought Paths Without Losing Accuracy