BlenderBuilder: Thinking from a builder’s perspective to synthesis more complex objects

3D coding provides a promising path toward editable, controllable, and executable 3D content generation. However, existing 3D coding agents predominantly construct objects from a parametric perspective, requiring models to describe complex geometry through code or combinations of simple primitives. This formulation becomes increasingly difficult for objects with complex surfaces, fine-grained structures, and long construction processes. We introduce BlenderBuilder, an operation-centric framework that reformulates 3D modeling from describing target geometry to reasoning about how to construct it. Given a text description or reference image, a Builder plans semantically meaningful modeling operations and selects appropriate tools from a structured Blender Toolkit, while a Coder translates these decisions into executable Blender code. An independent Critic actively inspects intermediate results from informative viewpoints and provides targeted revision feedback, forming an iterative construction and refinement loop. To systematically evaluate 3D coding at both object and scene levels, we further introduce BlenderObject-250 and BlenderScene-40, covering diverse objects and complex indoor and outdoor scenes. Extensive experiments show that BlenderBuilder achieves state-of-the-art performance on both object-level and scene-level generation, with larger gains on more complex scenes. Surprisingly, clearer construction planning and targeted revision also reduce redundant exploration and token consumption. Moreover, evolution experiments show that BlenderBuilder provides increasingly larger gains as the underlying foundation model becomes stronger, suggesting that operation-centric construction can continue to benefit from future advances in 3D reasoning and code generation.