Agentic Reasoning for Large Language Models