Truss structures must meet prescribed mechanical requirements, as well as practical limitations related to fabrication, transportation, and assembly. As a result, truss design is a highly constrained problem, where traditional optimization approaches are hindered by computational demands and slow convergence in large search spaces. In this talk, we propose a generative framework that reduces the computational cost of automated truss design by using grammar rules to embed engineering knowledge into the design process. We approach truss optimization through the lens of Markov decision processes, where actions involve adding or removing truss members to optimize a design objective. By constraining this sequential decision problem with grammar rules, we simulate progressive construction processes, where the final design is reached through intermediate feasible configurations. The application of these rules to each decision step reflects continuous human-computer interaction, narrowing the search space of alternative designs. The action policy is learned using Monte Carlo Tree Search (MCTS), a decision-time planning reinforcement learning algorithm. We initialize a seed configuration and use MCTS to iteratively navigate feasible designs, balancing exploitation and exploration. We demonstrate the superior design capabilities and computational efficiency of our approach compared to grammar-based Q-learning and deep Q-learning solutions. Furthermore, we highlight the robustness of our framework in progressive construction setups, where the seed configuration grows, mimicking an additive construction process.