aDSL: Agentic 3D Creation via Joint Agent-Program Design

A new agent-centric 3D creation stack tries to solve a familiar problem for LLM-driven content tools: they can describe intent well, but they often stumble when forced to emit exact geometry. The core idea is to co-design the language and the agents around the strengths of LLMs, rather than asking a general-purpose model to wrestle with brittle numeric interfaces.

The custom DSL, called aDSL, emphasizes composability and spatial reasoning. Instead of leaning on absolute coordinates, it lets agents express geometry through relationships and constraints, which should make generated content easier to control, edit, and interpret. That matters for teams building procedural tools, authoring assistants, or any pipeline where downstream edits are part of the workflow.

On top of the language sits a training-free multi-agent system that runs a Plan-Execute-Critic loop. One agent decomposes the request, another writes code, and a critic uses execution feedback to catch errors and constraint violations before they harden into broken assets. In practice, that kind of iterative repair is the difference between a demo and something artists or technical designers can actually iterate on.

The system reportedly outperforms prior LLM-based baselines on text-to-shape and image-to-shape tasks while keeping explicit structure and editability intact. It also extends to articulated object creation and structured scene composition, which makes it relevant beyond isolated meshes. The code is available, so teams interested in agentic content generation can start testing how far relational,...