BrickWright: Agentic Creation of Physically Feasible Brick Builds

BrickWright. Agentic Creation of Physically Feasible Brick Builds. By Ava Pun and nine others, from Carnegie Mellon University, Independent Researcher, and one other institution. Published on the twenty-second of October. Its code repository is a placeholder: a readme, and nothing to run yet.

Frontier coding agents are getting increasingly good at 3D modeling. Building something meant to be assembled in the physical world is a harder test. Here agents often produce invalid assemblies: parts that float, parts that collide, parts that never connect to anything. A model can look right on screen and still be impossible to put together.

The authors argue that agents do not need to be retrained for this. Giving them tools to assemble a build and to verify its physical feasibility is, they claim, sufficient for them to produce physically valid assemblies that still follow the prompt. They test that claim on LEGO brick builds with BrickWright, a framework that uses pre-trained agents to generate builds from text or image prompts.

Feasibility has a precise meaning here. A build must use only real parts, have no colliding or floating parts, and form a connected structure where that is possible. Those three conditions map directly onto the failures described above. Invented parts, collisions and floating pieces, and disconnected fragments are each ruled out by one of them.

BrickWright guides an agent to write code with BrickLayer, a Python toolkit with two main components. The first is an ergonomic brick assembly API, which halves inference cost compared with working in raw LDraw. The second is a validator that gives the agent feedback on physical feasibility, so problems are reported back to the agent rather than discovered later. The two halves answer the two parts of the authors' argument. The API is the tool for assembling, and the validator is the tool for verifying. Neither changes the agent itself; both change what the agent is working with.

Without this guidance, baseline agents reach physical validity rates from only nine percent, for GPT-6 Luna, up to sixty-nine percent, for GPT-6 Astra. Guided by BrickWright, agents reach one hundred percent validity on a sample of five hundred and twelve prompts. The gap between the weakest and strongest agents, which was wide without the tools, closes entirely on this measure.

Prompt adherence does not suffer for it. It improves for lightweight agents and holds steady for stronger ones. The guided agents also outperform specialized models on the text-to-brick task.

The claim is modest in machinery and large in effect: no new model, just an easier language to build in and a checker that says what is wrong. Validity goes from a matter of which agent you picked to something every agent reaches. Code is available from the project page.