Coarse-grained molecular dynamics (CGMD) enables simulation of large polymer systems with a substantial reduction in computational cost compared to atomistic molecular dynamics (MD), but remains limited by the laborious development of coarse-grained (CG) interaction potentials. Traditional potential parameterization, involving atomistic simulations, distribution extraction, and inversion techniques, demands expert involvement and extensive manual tuning, hindering the broader application of CGMD in polymer and materials research. Here, we present an automated CGMD pipeline powered by a multi-agent large language model (LLM) framework that decomposes the workflow into specialized agents for simulation requirement extraction, MD execution, bead mapping, potential generation, CGMD simulation, and analysis and reporting. Beginning with a high-level user query specifying polymer type and simulation goals, the system autonomously manages atomistic MD to generate bond, angle, and radial distribution functions, from which CG potentials are derived using Boltzmann inversion. These automated potentials are used to execute CGMD simulations, and standard validation criteria such as density, temperature, and energy convergence are evaluated. We demonstrate that this framework consistently produces CGMD outcomes that align with expected structural and thermodynamic trends, enabling users to obtain interpretable simulation results without specialized expertise. The multi-agent design of the system follows emerging strategies in autonomous scientific workflows, as exemplified by recent multi-agent AI approaches for materials design and simulation automation [1]. By automating the traditionally manual stages of CG potential development and simulation, this work reduces the workload on researcher, improves reproducibility, and lowers the technical barrier to deploying CGMD in polymer science and modeling. The proposed framework reshapes how CGMD is conducted, transforming CG potential development from a case-specific, expert-driven task into a reproducible and user-driven computational workflow.