Urban carbon flux estimation remains a critical challenge for net-zero planning, as conventional process-based models fail to capture the dynamic, multi-scale interactions between human decisions and physical systems. We propose the Spatio-Temporal Graph Multi-Agent Carbon Flux Model (STG-MACFM). This learning-driven framework replaces static emission inventories with a hierarchical agent architecture embedded within a GeoAI-enabled digital twin. The model decomposes the urban domain into numerous localised agents, each representing a building or neighbourhood unit. These agents maintain local states derived from real-time energy consumption, meteorological data, and building geometry, and they learn optimal intervention policies through proximal policy optimisation. A temporal attention mechanism within each agent captures delayed feedback loops, such as the thermal inertia of building materials or lagged occupant responses to policy changes. A global coordinator then constructs a dynamic urban topology graph using a graph attention network, where edge weights encode spatiotemporal dependencies like wind-driven CO₂ dispersion or shared grid constraints. This coordinator aggregates local states through a graph convolutional network to produce city-wide strategies, including carbon pricing rates or district heating setpoints. The global policy is optimised via a multi-agent deep deterministic policy gradient variant, with a reward function that penalises both total emissions and spatial inequity. The proposed model integrates seamlessly with existing data layers and simulation modules, replacing static emission factors with context-aware, real-time values. Furthermore, the closed-loop coupling between local agent behaviours and global strategic interventions enables the digital twin to simulate complex cross-scale socio-environmental interactions with high fidelity. This work introduces a novel paradigm for urban carbon modelling, where adaptive, graph-based coordination replaces rigid, top-down estimation, thereby offering a more realistic foundation for net-zero scenario optimisation.