Urban mobility systems are increasingly strained by dynamic societal demands, environmental pressures, and infrastructural limitations. Traditional simulation approaches—rooted in fixed rules and simplified behavioural assumptions—struggle to capture the complexity and variability of human travel decisions. This paper presents a novel framework that integrates Large Language Models (LLMs) into multi-agent mobility simulations to enable context-aware, adaptive, and intelligent decision-making in urban transport planning. Our approach leverages real-world survey data to generate diverse traveller profiles, embedding demographic, behavioural, and contextual factors into SimFleet, a cutting-edge mobility simulation platform. Unlike traditional rule-based models, our framework allows agents to learn, adapt, and refine their travel choices through iterative interactions with simulated urban environments. The framework, built on the SimFleet platform, is evaluated using diverse user profiles and varying levels of contextual information. Results show that LLM-driven agents can achieve high decision quality, behavioural consistency, and realistic adaptation over time. This work highlights the potential of combining LLM reasoning with multi-agent systems to support more intelligent, interpretable, and flexible simulation tools for urban mobility.