Toward agentic large language models in travel behavior research

Travel behavior research has traditionally relied on utility maximization frameworks and activity-based models that assume perfect rationality, yet struggle to capture the psychological, cultural, and contextual dimensions of human mobility decision-making. While large language models (LLMs) have been applied to transportation tasks, their use has largely been limited to regression and classification problems. This article proposes a paradigm shift toward deploying LLMs as agentic, simulated individuals that embody travelers’ roles and reasoning processes. We outline a unified agent framework to enable agentic LLMs to generate behaviorally rich, context-aware synthetic travel data. By reframing research objectives at the individual level rather than through aggregate metrics, this approach opens new avenues for exploring discretionary activity selection, intrapersonal household coordination, and the latent mechanisms underlying travel decisions. We discuss the architecture and potential applications of agentic LLM-based travel behavior models, while critically examining challenges related to alignment, bias, hallucination, domain suitability, and validation. Rather than replacing traditional models, agentic LLMs offer a complementary tool for bridging formal modeling with the lived experience of mobility, providing contextual explainability that illuminates not only where and how people travel, but why they choose to travel at all.