A Hugging Face team member demonstrates how the building-block economy applies to multimedia AI by chaining two Hugging Face Spaces via agents.md. The core idea: the hardest part of using state-of-the-art models is integration—SDKs, weights, GPUs, input formats, polling. To solve this, every Gradio Space now exposes a plain-text agents.md file that tells an agent exactly how to call it: schema URL, call/poll endpoints, file upload, and auth hint. An agent can read that file and drive the Space end-to-end without any client library or hardcoded integration. The real unlock is chaining: the output of one Space becomes the input to the next.

As a concrete example, the author asked a coding agent to build a 3D gallery of Paris monuments. The agent chained two Spaces: Ideogram (image generation) to produce clean, dark-background specimen shots of each monument, and TripoSplat (single-image to 3D Gaussian splat) to reconstruct a 3D model from each image. The agent then handled all the glue work: it noticed TripoSplat outputs are Y-down and flipped them upright, auto-framed each monument, compressed .ply files to .ksplat (~3x smaller) for fast loading, built a Three.js viewer with scroll-to-switch and drag-to-rotate UI, and deployed everything as a static Space. The only human inputs were taste-level adjustments like ‘make it zoomed out’. The author notes that several steps were the agent reacting to real-world constraints—a wide glass pyramid splats poorly, a thin obelisk is dull. This is exactly the outsourced R&D, fast iteration loop the building-block economy predicts.

Once the pipeline existed, spinning up new galleries cost just one sentence each: ‘Create a similar Space with splats for Japan’—and the agent produced six monument images, six splats, compression, viewer, and deployed Space per country. Galleries for Egypt and Japan were created with the same two Spaces and same agents.md, only the prompts changed. The marginal cost of a new multimedia app falls toward the cost of describing it.

Why this matters: models become composable primitives. A SOTA image model and a SOTA 3D reconstruction model from different organizations can be 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. The barrier to integrating multimedia AI was integration, and it’s largely gone. The post includes curl commands for readers to try with their own coding agents, and the full reproducible pipeline is available in the Space repo.