Why
Three-dimensional modelling was one of the crafts assumed to be safe from language models: spatial, visual, tacit, judged by eye. The demonstration removes two of those defences at once. The model does not need to be spatial if it can write code against a scene graph, and it does not need a human eye if it can render and look at the image itself. What remains is the third defence, taste, which the video does not test; the result is recognisably the Sagrada Família rather than a good one. The transferable lesson is about tool design: any application with a scriptable API and a way to observe results, a game engine, a CAD tool, a spreadsheet, a chain devnet, can be driven the same way, and the quality of the loop depends on how well the tool reports back.
How it works
The setup
Windows PowerShell running the GPT 6 Codex CLI, with a Blender MCP server started in the background so the model can call Blender operations as tools (00:53). The model is switched from the default medium setting to GPT 6 Astra High for stronger reasoning and execution (01:35). Hardware is an RTX 4070 laptop GPU (03:27), which matters because the model later checks it before choosing mesh budgets.
Planning and research without being asked
The user asks aloud for "Gaudí's cathedral in Spain" in Blender (02:05). The agent opens the Sagrada Família's official site, finds the drawings and interior photographs as PDF and reads them with a PDF skill (02:25). It checks system RAM and GPU (03:15), then produces a staged plan in about three minutes: block-out, exterior architecture, interior columns and ceiling, stained-glass materials, surroundings (city, park, lake), then optimisation and render inspection, using the completed 18-tower design as the reference (03:52). The user asks for a progress report every five minutes (04:51).
The self-correction loop
- Its own cameras. The agent places three exterior and two interior cameras so it can inspect the scene without a human (05:56).
- Structure in passes. Interior columns first, about 580 thousand faces, then stained glass, then the façades at about 1.9 million (06:17).
- Looking at the render. After the first exterior render it reads the image, sees that the roof junction under the central tower is empty, writes additional code and closes the gap (07:16). In the interior it judges the light too dark, adds lights and re-renders (09:06).
- Unrequested context. Spanish-style neighbouring buildings, a park, a lake and street trees are placed from web reference without being asked (09:42).
Result
Triangle count peaked at 3.17 million and was optimised to 2.78 million (11:52). The agent wrote a markdown summary of the work into the documents folder (11:34). Total time including Cycles-based inspection renders: 28 minutes 49 seconds, file about 164 MB (13:11), for a scene a person would spend days on. The video's conclusion: complex 3D form can be handled by an agent that researches, codes, renders, inspects and fixes in a loop through MCP tools (10:57).
| Step | Who did it before | Who did it here |
|---|---|---|
| Find reference drawings | the artist | the agent, via the official site and a PDF skill |
| Choose mesh budget | the artist, by feel | the agent, after checking RAM and GPU |
| Build geometry | the artist, by hand | the agent, as Blender code through MCP |
| Inspect the result | the artist's eye | the agent's own cameras and image reading |
| Fix defects | the artist | the agent, in a second pass |
Where it lands in Jayverse
- Game: the Unity street can be generated, not modelled. The 3D street where Verex markets appear on boards needs buildings, signage and props; a Blender MCP loop that produces low-poly assets to a stated triangle budget, renders them from fixed cameras and exports glTF for Unity is a weekend experiment with a clear evaluation: does it load, does it fit the budget, does it read as a street.
- Tech #62's boundary file applies to tools too. The agent succeeded because Blender exposed a clean scripting surface and a render it could look at. Any Jayverse tool we hand to an agent, the Foundry scripts, the devnet RPC, the Verex CLI, should offer the same two things: a scriptable interface and a way to observe the outcome.
- Auditor: "inspected by the agent" is a claim to record. The self-check rendered five cameras and read the images; that is an evaluation step, and its rule (which cameras, which defects it looks for) belongs in the record next to the artefact, or the fix at 07:16 is luck.
- Eng: this is the demo to describe in an interview. "An agent drove Blender through MCP, inspected its own renders and fixed a defect" is a concrete answer to "how do you use AI", better than any list of tools.
Verified and unverified
Verified on 2026-09-19: Blender exposes a Python API and community MCP servers for it exist; the Codex CLI supports MCP tool servers; the Sagrada Família's design has eighteen towers when complete. Taken from the video and not independently checked: the model name and setting ("GPT 6 Astra High"), all mesh counts, the 28-minute-49-second total, the 164 MB file size, and the specific defects the agent found. The video shows one run; how often the loop converges without human help is not measured.
Sources: AZTechnology — GPT 6 x 블렌더 3D = 이제는 바이브 모델링 시대, YouTube · related items: Tech #62 (agentic engineering writes the boundaries), Tech (AI engineer builds the car), Jayverse #5 Game.
Key expressions
| Expression | 뜻 · 쓰이는 자리 |
|---|---|
| vibe modeling | 바이브 모델링(자연어로 말해 에이전트가 3D 모델을 만들게 하는 방식) · vibe coding에서 온 신조어. "what it calls vibe modeling" |
| MCP | Model Context Protocol(모델이 외부 도구·데이터를 표준 방식으로 호출하게 하는 프로토콜) · 에이전트 도구 연동의 표준 이름. "connected to Blender through an MCP server" |
| scene graph | 씬 그래프(3D 장면의 객체를 트리로 조직한 자료구조) · 3D 엔진·Blender 내부 모델. "write code against a scene graph" |
| block-out | 블록아웃(큰 덩어리로 형태를 잡는 첫 단계) · 3D·게임 아트 용어. "block-out, exterior architecture" |
| façade | 파사드(건물 정면) · 건축 용어, 영어에서도 프랑스어 철자 유지. "then the façades at about 1.9 million" |
| stained glass | 스테인드글라스 · 재질(material) 설정의 대상으로. "stained-glass materials" |
| mesh / triangle count | 메시 / 삼각형 수(3D 모델의 복잡도 척도) · 성능 예산을 말할 때. "a 2.78-million-triangle scene" |
| render (v./n.) | 렌더하다 / 렌더(3D 장면을 이미지로 계산해 내기) · 검수 대상. "renders, notices a hole" |
| Cycles | Blender의 경로 추적 렌더러 이름 · 고품질·느린 렌더. "Cycles-based inspection renders" |
| self-correction loop | 자가 수정 루프(만들고, 보고, 고치기를 반복) · 에이전트 설계의 핵심 패턴. "The self-correction loop" |
| tacit | 암묵적인(말로 옮기기 어려운 기술) · 장인의 기술을 설명할 때. "spatial, visual, tacit, judged by eye" |
| proof of concept (PoC) | 개념 증명(될 수 있음을 보이는 작은 시연) · 이 사이트의 항목 타입 이름과 같은 말. "a proof of concept for the Game service" |
| low-poly | 로우폴리(폴리곤이 적은, 가벼운 모델) · 게임 자산 요구 조건. "low-poly assets to a stated triangle budget" |
| glTF | GL Transmission Format(3D 자산 교환용 표준 파일 형식) · Blender → Unity 내보내기 형식. "exports glTF for Unity" |
| scriptable surface | 스크립트로 조작 가능한 표면(API) · 도구를 에이전트에 열어 주는 조건. "a clean, scriptable surface" |
| observe the outcome | 결과를 관찰하다 · 피드백 루프의 두 번째 조건. "a way to observe the outcome" |
| converge | 수렴하다(반복이 목표에 도달하다) · 루프가 끝나는지를 말할 때. "how often the loop converges without human help" |
| taste | 취향(좋고 나쁨을 가르는 판단) · 자동화가 남긴 마지막 방어벽으로. "the third defence, taste" |