CMU-Stanford BrickWright Lifts Agent LEGO Validity to 100%

A new framework from Carnegie Mellon and Stanford gets pre-trained coding agents to produce LEGO builds that actually fit together, lifting physical-validity scores from as low as 9% to 100% on 512 text and image prompts.

The system, BrickWright, is described in a paper posted to arXiv on Friday by Kevin Dai and colleagues at CMU with Maneesh Agrawala at Stanford. It hands agents a Python toolkit called BrickLayer that pairs an ergonomic brick-assembly API with a validator checking for colliding, floating, or disconnected parts. Baseline agents without BrickWright land at physical-validity rates "ranging from only 9% (GPT-6 Luna) to 69% (GPT-6 Astra)," the authors write; with the framework, "all agents...achieve a 100% validity score without a significant cost increase or drop in fidelity," including lightweight models like GPT-6 Luna and Claude Haiku 5.5.

The cost argument is the quieter half of the finding. The BrickLayer API "halves inference cost compared to working in raw LDraw," and lightweight agents guided by BrickWright hit full validity "at 2% the price of frontier models." The paper also notes that GPT-6 Astra with Codex, when asked to build LEGO freehand, already "writes scripts to generate and validate its output more than 80% of the time," which suggests the toolmaking instinct is there in strong agents but incomplete without a domain-specific validator.

BrickWright arrives in the middle of a dense run of agent-tooling research we have been tracking in our agents feed, which has logged 523 stories over the past 90 days. Code is posted at the project page.