<p>Automating the optimal design of welded plate girders offers significant cost and time savings in structural engineering. This paper introduces a novel framework that combines multi-agent reinforcement learning (MARL) with an inference chain and a shrink refinement procedure to minimize the manufacturing costs of plate girders while adhering to Eurocode constraints. The optimization problem for plate girders is defined using seven discrete design variables and a five-component cost model. A cooperative MARL architecture, known as Comm-DDQN, is trained to sequentially select girder dimensions while ensuring structural feasibility. During inference, a two-stage search process—consisting of beam search followed by stochastic sampling—utilizes the learned policy to efficiently explore high-quality design candidates. A subsequent UF-aware local search (shrink refine) fine-tunes the final design to reduce over-conservatism, pushing the utilization factor (UF) close to 1.0 without violating constraints. We evaluate this approach across 100 diverse load cases and compare it with a Genetic Algorithm (GA) baseline. The proposed MARL method achieves cost outcomes comparable to GA while reducing per-design optimization time considerably. Parametric analysis of the beam search width (W), number of stochastic samples (N), and refinement toggle shows that a broader search and refined approach consistently enhance cost efficiency, reducing average costs by approximately 5–7% and producing designs with UF ≈ 1.0, albeit with a slight increase in inference time. These findings demonstrate that a sequential MARL approach, combined with a beam-stochastic inference chain and refinement, can effectively automate steel girder design, balancing cost efficiency, structural safety, and computational performance. The paper also discusses potential extensions to more complex design scenarios and real-time optimization in design offices.</p>