Vibe-Blender: A Blender-Native Agent Framework for Iterative 3D Design

BlenderAgent is a Blender-native agent framework for iterative 3D design. Existing research has advanced text-to-3D generation, CAD generation, and general coding agents, yet offers limited support for grounded, stepwise design work inside a live 3D authoring environment. BlenderAgent addresses this gap by organizing design as a cycle of code writing, execution, rendering, inspection, and revision within Blender. The system combines a coding-agent harness with Blender-aware tools, including live scene context, visual feedback, and traceable interaction history, so model actions remain inspectable and tied to the evolving scene. This thesis examines how large language models can be integrated into real 3D design workflows, what distinguishes 3D design from text and code generation, and why human guidance remains important in design tasks. Evaluation includes BlenderBench benchmark results and qualitative case studies of iterative scene editing. The findings suggest that code-agent-style interaction is a promising path for AI-assisted 3D design, especially when paired with render-based feedback and native software context. The results also reveal continuing limits in spatial reasoning and long-horizon planning, supporting a human-in-the-loop view of AI-assisted design. The prototype is open-sourced at https://github.com/veoery/BlenderAgent/tree/live-session.