Agentic AI for Argument Extraction: A Comparative Study of Prompting, ReAct, and Multi-Agent Systems

Abstract

In theory, extracting arguments from news articles should be simple, but in practice it proves to be more difficult. This paper explores three approaches to address this problem: classical static prompting, a single adaptive agent using ReAct, and a multi-agent team of specialized reasoners. The results were striking: multi-agents achieved 41% accuracy, seven percentage points better than the conventional prompting baseline, which achieved about 34%. Both multi-agent and ReAct produced usable results for every article, whereas static prompting produced empty or garbled output in roughly 10% of cases. The advantage of multi-agent systems was greater for difficult argument categories, particularly counterarguments and evidence. These findings suggest that adaptive and collaborative systems can improve both the coverage and reliability of argument extraction for applications such as fact-checking, policy analysis, and critical-reading tools.