Large Language Models (LLMs) have become ubiquitous in day-to-day life. Recent advancements concerning the integration of LLMs, multi-agent architectures and embedded systems have enabled the deployment of complex, data-oriented applications for Smart Cities. While contemporary urban mobility monitoring heavily relies on Floating Car Data (FCD) to capture spatial traffic patterns, such approaches operate reactively and lack human-centric context. Conversely, critical urban disruptions (accidents, scheduled events, flash floods, roadworks, etc.) are often reported by citizens via heterogeneous text and audio streams such as social media, news feed and radio broadcasts, sometimes even before they manifest in physical traffic flows. Existing semantic enrichment methods, however, focus primarily on historical data and rarely exploit foundation models or agentic frameworks for real-time streaming context. This work presents a multi-agent approach that continuously collects semantic urban data from social media, Google RSS feeds  and live radio broadcasts, transcribes and extracts structured events, and aligns them with vehicular trajectories to produce enriched, explainable routes. The pipeline is designed as a Mixture-of-Experts (MoE) system communicating through versioned Pydantic contracts. It is coordinated by an orchestrator agent alongside a dedicated reasoning agent operating under the ReAct paradigm, which includes a human-in-the-loop validation protocol. We instantiate and evaluate the system using open and citizen-generated data from Belo Horizonte, Brazil, during an uninterrupted 24-hour operational window. Hence, our results demonstrate the system’s efficacy regarding extraction quality, faithfulness, and operational cost-efficiency, while generating explainable insights into traffic dynamics from a human-centric perspective to enable context-aware mobility forecasting.