A 2,175-piece pagoda is a striking demonstration of an AI agent controlling a CAD workflow. It is not yet evidence that someone can build the design with LEGO bricks. That distinction separates LDraw Nova's expressive digital models from the harder problem its developer has not claimed to solve: producing an assembly that survives real-world construction and handling.
Carlos Antelo's open-source toolkit lets general-purpose agents write programs that generate LDraw models. In the October 4 account, its Sakura Garden combines a five-story pagoda, cherry trees, a torii gate and a koi pond. Claude Opus 5.5 produced it from an open-ended request to imagine the most beautiful model possible. The reported 2,175 pieces are digital catalog parts, not a completed physical set.
The agent edits a program, not thousands of coordinates
LDraw Nova is neither a specially trained LEGO model nor simply an image generator. The agent reads instructions and examples, prepares a JSON plan for the model and its submodels, then writes Python using higher-level construction primitives and reusable patterns. The program emits LDraw files, a headless renderer produces views, and a multimodal agent examines the result before revising the program.
That abstraction matters. Raw LDraw describes part identifiers, colors, positions and orientation transformations. Antelo's earlier attempts to generate long lists of coordinates and rotation matrices directly were unsuccessful. Reusable code gives the agent a more manageable way to change a roof, repeat a wall or adjust an entire subassembly, rather than place thousands of pieces through individual commands.
The toolkit includes part and example search, collision and gap checks, rendering and a Docker-based web application. It supports OpenAI, Anthropic and OpenRouter providers. Optional semantic reranking requires an API; full-text search remains a fallback. The underlying LDraw library is a community-maintained, unofficial resource, not a LEGO Group-sponsored product. The source describes 16,873 unique shapes and patterned parts after its June 2026 update; that is a dated catalog count, not a claim about today's inventory.
The project repository preserves more than pictures: editable Python and LDraw files, rendered views, an interactive viewer, glTF export for Blender, and the agent's conversation and reasoning. Its gallery also includes a cathedral and Tidal Observatory attributed to GPT-6 Astra, Copper Bean apartments from an earlier Claude Opus, and an unfinished Atlas Crane from Opus 5.5. These are examples of different generated designs, not a controlled ranking of providers.
A valid part identifier does not prove a sound building
A catalog-backed design has an advantage over an invented image: its pieces can be identified. But an existing part identifier does not establish that a particular color is available, that enough units can be sourced, or that an older part is still sold. Purchasing and inventory remain separate checks.
More importantly, Antelo acknowledges that the system lacks physics validation, and the showcased models had not been physically assembled in the reported account. Collision and gap checks can catch some geometric errors. They do not establish clutch strength, load paths, balance, resistance to torsion or whether an overhanging roof and tree can tolerate handling.
Construction order adds another constraint. A final shape may appear stable while an intermediate stage falls apart or requires access that later pieces block. A useful build plan must account for actual connections, assembly access and the sequence in which support becomes available, not merely display a completed object from attractive angles.
Scale can magnify these weaknesses. A flawed helper routine may replicate the same bad connection hundreds of times. Rendered surfaces reveal appearance, not forces; a higher piece count increases the amount of inventory and structural checking rather than making it unnecessary.
LegoGPT sets a different evidentiary benchmark
The LegoGPT research introduced by Carnegie Mellon collaborators in 2025 tackles a more constrained task. Instead of predicting the next word, its model predicts the next brick. StableText2Lego contains more than 47,000 structures representing over 28,000 unique 3D objects, paired with descriptions and assessed for stability. Its voxel-derived representation covers 21 common object categories.
Those restrictions reduce expressive freedom compared with Nova's ambitious scenes, but they make physical validity more central to generation. LegoGPT uses validity checks and physics-aware rollback: invalid placements are rejected, and generation returns to a preceding valid state. The study's notion of stability includes no floating or collapsing pieces and strong structural integrity; buildability also involves standard parts and sequential human or robot assembly.
Selected outputs were assembled manually and with a bimanual robot. That is stronger physical evidence than a render. It is not a guarantee about every future output, nor a head-to-head demonstration that the system can construct Nova's 2,175-piece garden. The source also associates later discussion with the BrickGPT name; that should not be read as proof of an equivalent new release or benchmark.
Nova's files can be opened in tools such as LeoCAD, which supports LDR and MPD, submodels and multistep instructions. Format support is useful for editing and inspection. It does not mean Nova has already generated and validated complete, physically executable instructions.
Physics feedback must change the design
The promising next step is not simply attaching a “stable” score to a finished render. Structural feedback would need to change the agent's plan, part choices, subassemblies and construction order. Connection graphs, center-of-mass estimates and identification of weak joints could help, but these are possible extensions, not capabilities demonstrated by Nova in the account.
Even a sophisticated simulation simplifies real clutch forces, friction, manufacturing tolerances and handling. A physical prototype remains decisive. Testing a pagoda tier, roof corner, tree or crane joint would expose different failure modes without committing immediately to a large purchase.
Antelo considered 3D-printing a smaller model. That had not been reported as a completed test, and a printed miniature would not validate the connections of a LEGO assembly. A useful physical demonstration would publish the original CAD, bill of materials, instructions, build time, manual corrections and failures, then show the model being handled normally.
What progress would look like
Three signals would move this beyond impressive visualization: a documented real-brick build; evidence that failed structural tests caused the agent to revise its program; and repeated trials across providers and prompts using a fixed parts library and common criteria. Completion time, invalid placements, manual intervention, sourcing conflicts and physical failures would make those trials more informative than piece count alone.
Minifigures, Technic machinery and engines are roadmap ideas rather than established capabilities. Axles, pins, moving joints, alignment and torque would bring additional constraints. The project has VR viewing, but Antelo reported performance and usability problems; immersion does not replace mechanical or assembly testing.
LDraw Nova already illustrates a useful software pattern: let an agent operate a structured tool through reusable code, then inspect and revise its output. For builders, however, its files remain editable starting points rather than certified kits. The next milestone is not a more elaborate screenshot, but an assembly whose parts, sequence and structure can be checked.
Sources: October 4 report; original reporting aggregation entry.