Abstract: Rapid and accurate assessment of earthquake losses provides an important basis for post-earthquake emergency response and decision-making. Conventional loss assessment workflows generally rely on manual collection and processing of data on seismic source parameters, ground motion, population, and building exposure. Such workflows face difficulties in integrating heterogeneous data from multiple sources, often providing insufficient spatial resolution, and may not meet time-critical assessment requirements. To support post-earthquake emergency responses, in this study, a rapid earthquake loss assessment workflow orchestrated by a large language model (LLM) agent was developed. The workflow does not alter existing models for ground-motion prediction, exposure modeling, or structural fragility analysis. Instead, it automates the links between task decomposition, data retrieval, parameter configuration, model invocation, and result aggregation, thereby improving the consistency, traceability, and reproducibility of multisource data processing and loss estimation. This framework integrates spatial data on topography, site V_S30, population, and building asset exposure. It uses GEM building exposure data, OpenQuake fragility curves, and a spatially correlated loss-sampling method based on β marginal distributions and a Gaussian Copula to propagate the uncertainty in building damage and loss estimates. The April 2026 M_W7.4 Molucca Sea earthquake in Indonesia was used as a case study. The assessment was limited to direct structural losses to residential buildings caused by ground shaking and the risk of fatalities resulting from structural damage. It excluded impacts from tsunamis, landslides, and fires as well as infrastructure and business interruption losses. The results show that the framework can automatically complete the entire assessment process, from reading seismic source parameters to generating probability distributions of economic loss and number of fatalities. The expected direct structural loss to residential buildings was approximately USD 3.6282 million. The mode of the fatality distribution was zero, and the number of fatalities was highly likely to remain below five. Compared with the USGS PAGER estimates and publicly available earthquake damage reports, the estimated economic losses and fatality levels were broadly of the same order of magnitude. Validation against historical events in the same region showed good agreement. This study provides a reusable, workflow-based technical approach for rapid earthquake loss assessment assisted by a large language model agent.