AI代理如何串聯兩個Hugging Face Spaces，打造3D巴黎畫廊

Back to Articles How an Agent Built a 3D Paris Gallery by Chaining Two Hugging Face Spaces Community Article Published June 9, 2026 Upvote 23 +17 Mishig Davaadorj mishig Follow An agent built a 3D Paris gallery from two Hugging Face Spaces. I asked a coding agent to build a beautiful website showcasing the monuments of Paris as 3D Gaussian splats. I never opened an image generator. I never touched a 3D reconstruction tool. The agent produced every asset (the images and the 3D splats) by calling two Hugging Face Spaces directly, then wired them into a cinematic viewer. Here's the result, live as a static Space: 👉 mishig/monuments-de-paris This post is about how that's possible now, and why I think it's a preview of how a lot of multimedia software gets built from here on. The building-block economy comes for multimedia Mitchell Hashimoto recently described a shift he calls the building block economy: the most effective path to software is no longer a polished monolith, but small, well-documented components that others (increasingly agents) can assemble. His key observation: AI is okay at building everything from scratch, but it is really good at gluing together proven pieces. That thesis has mostly been told with code libraries. But the same forces are hitting multimedia AI. The hard part of using a state-of-the-art image model, a video model, a TTS model, or a 3D reconstruction model was never the model. It was the integration: SDKs, weights, GPUs, input formats, polling. If each model were instead a documented, callable block, an agent could glue them together the same way it globs together npm packages. That's exactly what Hugging Face Spaces have quietly become. Every Space is a building block, via agents.md The Hub hosts thousands of state-of-the-art models (a huge share of them open-weights), and most are deployed as interactive Spaces. As of now, every Gradio Space also exposes a plain-text agents.md that tells an agent exactly how to call it: curl https://huggingface.co/spaces/VAST-AI/TripoSplat/agents.md returns everything needed in one shot: the schema URL, the call and poll templates, how to upload files, and the auth hint: API schema: GET .../gradio_api/info Call endpoint: POST .../gradio_api/call/v2/{endpoint} {"param_name": value, ...} Poll result: GET .../gradio_api/call/{endpoint}/{event_id} File inputs: POST .../gradio_api/upload -F "[email protected]" Auth: Bearer $HF_TOKEN No client library. No hardcoded integration. An agent reads that, and it can drive the Space end to end. Set an HF_TOKEN and you're going. You can find these instructions on any Gradio Space via its Agents button: The real unlock is chaining: the output of one Space becomes the input to the next. Prompt → image → 3D. That's the whole pipeline behind this gallery. The worked example: Paris monuments → splats The agent chained two Spaces: Image: an image-generation Space turned each monument into a clean, dark-background "specimen" shot (and the Eiffel Tower into a little diorama on a plinth). Prompt in, image out. Splat: VAST-AI/TripoSplat reconstructed a 3D Gaussian splat (.ply) from each single image. Image in, 3D out. Generated image Reconstructed splat The six source images the agent generated, all isolated on black, ready for single-image 3D reconstruction: From there the agent did the "glue" work too. It noticed TripoSplat outputs are Y-down and flipped them upright, auto-framed each monument, compressed the .ply files to .ksplat (~3× smaller, so they load fast), built a Three.js viewer with a scroll-to-switch and drag-to-rotate UI, and deployed the whole thing as a static Space. The only human inputs were taste-level: "make it zoomed out," "replace the obelisk with something better for splatting," "the transition lingers too long." Several of those steps were the agent reacting to reality. A wide glass pyramid splats poorly. A thin obelisk is dull. A single-view reconstruction infers the back. That is exactly the "outsourced R&D, fast iteration" loop the building-block economy predicts, except the R&D was a conversation. Two prompts, a whole new gallery The real test of a building block is how cheaply you can reuse it. Once this pipeline existed, spinning up entirely new galleries cost about one sentence each. "Create a similar Space with splats for Japan," then the same for Egypt, and the agent did the rest: six monument images, six splats, compression, a viewer, and a deployed Space, per country. 🏛️ Monuments of Egypt: the Great Pyramid, the Sphinx, Abu Simbel, the mask of Tutankhamun, Karnak, the Colossi of Memnon. <video autoplay loop muted playsinline width="100%" src=""> ⛩️ Monuments of Japan: Tokyo Tower, Himeji Castle, Kinkaku-ji, Osaka Castle, the Great Buddha of Kamakura, the Itsukushima torii. <video autoplay loop muted playsinline width="100%" src=""> Same two Spaces, same agents.md, only the prompts changed. That is the building-block economy in one line: the marginal cost of a new multimedia app falls toward the cost of describing it. Why this matters Models become composable. A SOTA splat model and a SOTA image model, from different orgs, chained with zero integration code. The Hub's open-weights catalog turns into a library of callable multimedia primitives. Agents prefer what's documented and reachable. agents.md makes a Space trivially reachable, so an agent will pick it over a model it has to set up by hand. That is the same dynamic Hashimoto flags for open-source libraries. The barrier was integration, and it's largely gone. "Turn a prompt into a rotating 3D monument" used to be a project. Here it was a step in a pipeline. Try it yourself Point your own agent at a Space's agents.md and let it cook: # image generation curl https://huggingface.co/spaces/ideogram-ai/ideogram4/agents.md # single-image to 3D gaussian splat curl https://huggingface.co/spaces/VAST-AI/TripoSplat/agents.md Paste either link into your coding agent (Claude Code, etc.), set your HF_TOKEN, and ask it to build something. The full, reproducible pipeline for this gallery, the scripts that hit those two agents.md endpoints, lives in the Space repo. The building blocks are sitting right there on the Hub. The agents already know how to glue. I asked a coding agent to build a beautiful website showcasing the monuments of Paris as 3D Gaussian splats. I never opened an image generator. I never touched a 3D reconstruction tool. The agent produced every asset (the images and the 3D splats) by calling two Hugging Face Spaces directly, then wired them into a cinematic viewer. Here's the result, live as a static Space: 👉 mishig/monuments-de-paris This post is about how that's possible now, and why I think it's a preview of how a lot of multimedia software gets built from here on. The building-block economy comes for multimedia Mitchell Hashimoto recently described a shift he calls the building block economy: the most effective path to software is no longer a polished monolith, but small, well-documented components that others (increasingly agents) can assemble. His key observation: AI is okay at building everything from scratch, but it is really good at gluing together proven pieces. That thesis has mostly been told with code libraries. But the same forces are hitting multimedia AI. The hard part of using a state-of-the-art image model, a video model, a TTS model, or a 3D reconstruction model was never the model. It was the integration: SDKs, weights, GPUs, input formats, polling. If each model were instead a documented, callable block, an agent could glue them together the same way it globs together npm packages. That's exactly what Hugging Face Spaces have quietly become. Every Space is a building block, via agents.md The Hub hosts thousands of state-of-the-art models (a huge share of them open-weights), and most are deployed as interactive Spaces. As of now, every Gradio Space also exposes a plain-text agents.md that tells an agent exactly how to call it: curl https://hugging