Large language models (LLM) have shown strong capabilities across a wide range of tasks, yet they can still produce hallucinated content and exhibit inconsistent reasoning. These shortcomings pose challenges in scenarios where dependable and accurate outputs are essential. Addressing this problem, this research presents the design, implementation and evaluation of a collaborative multi-agent (MA) framework that enhances the accuracy and reliability of LLM outputs. The proposed framework employs multiple LLM agents working in parallel to generate candidate responses for a given task. A novel aggregation algorithm is introduced, utilizing a hybrid, unified uncertainty quantification (UQ) method to estimate the confidence levels of each agent’s output. These confidence scores are then used by an aggregator model to determine and synthesize the most reliable final output. Evaluation results demonstrate that the proposed MA framework outperforms the leading state-of-the-art (SOTA) MA architecture by 6% in general domain tasks, achieving 65% accuracy on the TruthfulQA dataset and improves performance by 6% on programming benchmarks, achieving a 75% Pass@1 score on the HumanEval dataset.