From Traffic Indices to Agentic AI: A Reference Framework for Urban Mobility in Megacities

Megacity mobility research has long relied on aggregate statistical indicators (composite traffic-planning indices andaccessibility surfaces) that capture the system at rest but are disconnected from the operational decisions shaping mobilityminute to minute. In parallel, Agentic AI offers a reference paradigm of semantically interoperable autonomous agents that negotiate and coordinate across open networks. We propose a unifying Agentic AI reference architecture for urbantransportation that maps any composite traffic index and any accessibility surface onto agent utility functions, negotiation-protocol primitives, and shared semantic ontologies. The architecture is instantiated in simulation only: no live agent-to-agent endpoints, no runtime large language models, no cross-organisation interoperability experiment; these are explicitly listed as next steps. A 12-borough London case study, evaluated over $N=30$ seeds on an origin-destination matrix calibrated to the 2021 UK Census commuter-flow aggregates, benchmarks four regimes (historic, adaptive, MaxPressure, and our Agentic policy) across four scenarios covering equity, corridor prioritisation, incident response, and their combination. The Agentic regime reduces the accessibility-deficit Gini coefficient by \textbf{23--58\,\%}, travel-time coefficient of variation by up to \textbf{41\,\%}, and mean travel time by \textbf{4--9\,\%} relative to the historic baseline; in the joint equity-plus-incident scenario it attains the per-column best within $95\,\%$ confidence on travel-time coefficient of variation and mean travel time while improving on every metric over the historic baseline. Amicroscopic testbed in the SUMO simulator on a $4\!\times\!4$ signalised grid ($120$~runs across three demand regimes and five peripheral-boost values) traces an explicit equity--efficiency Pareto frontier; in the saturation and over-saturation regimes the agentic policy matches or beats a SCOOT-style adaptive controller on mean travel time and throughput at every boost level, with travel-time variance reduced by up to a third.