Webinar link.

Decision support tools become agentic when they exhibit a degree of agency—the capacity for contextual reasoning, goal-driven actions, social engagement, self-reflection, and external resource utilization—partially enabled by rapid advances in multimodal large language models. Such agency creates substantial potential to improve decision quality across the transport system lifecycle, from long-range planning to real-time operations; however, it also introduces severe challenges. Siloed, uncoordinated AI agents can produce unstable, opaque, and difficult-to-govern behaviors, raising concerns about system-level coherence, accountability, and alignment with societal objectives. This talk overviews the framework of agentic transport systems (AgTS) and suggests that the rise of agentic AI makes rigorous modeling of system dynamics even more indispensable. Partially Observable Markov Decision Processes (POMDPs) provide a principled foundation for modeling sequential decision-making under uncertainty. However, their standard formulation is insufficient to systematically evaluate and optimize agentic systems in complex socio-technical environments, as it assumes fixed reward functions and predefined state–action spaces. This presentation outlines principled extensions of POMDPs to endogenize reward specification and capture context-dependent, tool-augmented action spaces, positioning them as a unifying—yet evolving—analytical backbone for AgTS.