Generating high-quality, comprehensive arguments remains a key challenge in NLP. While LLMs are powerful text generators, relying on a single LLM often reveals limitations in achieving the structured reasoning and iterative refinement essential for robust, multifaceted argumentation. This paper introduces GuidedArg, a novel multi-agent collaborative framework designed to significantly enhance argument quality and comprehensiveness. GuidedArg employs four specialized agents for argument generation, quality evaluation, guiding question generation, and flow control, operating within an “Evaluate-Guide-Generate” iterative process. In this process, fine-grained, multi-dimensional feedback from the evaluation agent drives the creation of targeted guiding questions. These questions then direct the argument generation agent’s iterative revisions, encouraging more thorough analysis and exploration. Experimental results validate the framework’s effectiveness: the evaluation agent aligns strongly with human judgments and correlates positively with public acceptance rates; stance verification experiments demonstrate the system’s ability to discern quality differences arising from stance correctness. Crucially, end-to-end evaluations reveal significant improvements in average argument quality, including enhanced comprehensiveness, through this iterative optimization, with most samples converging efficiently. This showcases the efficacy of the guiding question mechanism and the overall GuidedArg framework.