The article introduces TRUST, a new reinforcement learning approach that improves decision-making in large language model-based agents by incorporating uncertainty quantification into reward design. This method enhances tool-use performance and reliability, addressing common issues of overconfidence and error accumulation in multi-step interactions.