UrbanAgent: Tool-Augmented AI for Cross-System Urban Tasks

A new tool-augmented agent framework, UrbanAgent, has been developed to address the fragmentation of digital services in modern cities and improve residents' ability to complete complex urban tasks. This framework integrates the reasoning capabilities of large language models with a comprehensive tool-set, enabling code execution, API calls, and adherence to the Model Context Protocol. UrbanAgent operates through an adaptive closed loop, allowing it to clarify ambiguous information, ground its tool usage in real-time observations, and ensure its final responses align with both observed evidence and task constraints. To facilitate rigorous evaluation, the researchers also introduced Urban-Eval, a benchmark specifically designed for cross-system urban requests, assessing both task outcomes and execution quality. On this benchmark, UrbanAgent achieved a 71% task success rate, outperforming strong baselines by 10 percentage points across various LLMs.