I gave a coding agent exactly two links and a vague idea: build a 3D gallery of Paris monuments. I never opened an image generator, never touched a 3D reconstruction tool. The agent produced every asset by directly calling two Hugging Face Spaces, then wired them into a cinematic viewer. The whole thing is live now as a static page.
The secret is a quiet infrastructure shift on Hugging Face. Every Gradio Space now exposes a plain-text agents.md file—essentially an instruction manual that tells an agent precisely how to call an API, upload files, and poll for results. No client library, no hardcoded integration. It’s a machine-readable contract. The agent read the manual for VAST-AI’s TripoSplat, sent it images, and got back 3D Gaussian splats. It did the same to generate the source images in the first place.
This is a working demo of what Mitchell Hashimoto calls the “building block economy.” AI is decent at writing code from scratch, but it’s exceptional at gluing together proven, documented components. The agent chained two Spaces—image generation to 3D reconstruction—and then did the unglamorous glue work. It noticed the 3D output was Y-down and flipped it upright, auto-framed each monument, compressed the files to .ksplat format for fast loading, and built a Three.js viewer with drag-to-rotate controls. The only human input was taste-level feedback: “make it zoomed out,” “the transition lingers too long.”
The real test of a building block is how cheap reuse becomes. Once the Paris pipeline existed, I asked for Japan. Then Egypt. Each new gallery cost roughly one sentence. The agent generated six new monument images, six new splats, compressed everything, and deployed a new Space per country. The marginal cost of a new multimedia app is collapsing toward the cost of describing it. The agents.md protocol makes a Space trivially reachable, which means an agent will pick it over a model it has to set up by hand. The building blocks are sitting on the Hub, and the agents already know how to glue.
