<p>Urban walkability is a critical determinant of health, safety, sustainability, and city life in general, yet most existing indices remain limited to amenity proximity and neglect the perceptual and morphological qualities that shape the walking experience. This paper proposes a Graph Machine Learning-centered multi-agent framework that integrates Large Language Models (LLMs), computational analysis, and human feedback to design and evaluate urban interventions. The framework positions Graph Machine Learning (GML) as the central predictive engine, capable of modeling network relationships and testing hypothetical scenarios, while LLMs act as perception interpreters that translate visual and textual information into experiential insights. Human agents validate and contextualize these results, ensuring alignment with policy and lived experience. A new Street Walkability Index (SWI) is introduced, combining traditional Walkscore metrics based on land-use and perception-derived data, to provide a multidimensional measure of walkability. Applied to Mexico City’s historic center, the system demonstrates improved predictive accuracy and interpretability compared to conventional models. Ablation studies confirm that integrating perceptual and topological features enhances performance, while intervention modeling shows the framework’s ability to simulate and evaluate interventions such as building massing and architectural program change. These results suggest that multi-agent GML systems offer a powerful decision-support approach for participatory urban evaluation, bridging data-based, perceptual, and human intelligences toward more equitable and actionable urban design strategies.</p>