How an Agent Built a 3D Paris Gallery by Chaining Two Hugging Face Spaces

Source: HuggingFace Date: 2026-06-09 URL: https://huggingface.co/blog/mishig/spaces-agents-md

Summary

An agent autonomously built an interactive 3D gallery of Paris monuments by chaining two HuggingFace Spaces: Ideogram 4 (text-to-image) feeding into TripoSplat (image-to-3D Gaussian splat). The key enabling mechanism is agents.md — a plain-text file that each Space exposes, describing its schema URL, call/poll endpoints, file upload procedures, and auth hints, giving an agent enough structure to drive the Space without a client library or hardcoded integration. The agent also handled coordinate corrections, compression (.ply → .ksplat, ~3× reduction), and Three.js viewer deployment, then replicated the pattern to produce Egypt and Japan galleries with minimal re-prompting.

Implications

Feeds the agentic-engineering patterns thread directly. agents.md is a lightweight convention — essentially a capability manifest for callable components — and if it gets traction across HuggingFace Spaces it becomes a de-facto standard for how agents discover and invoke multimedia pipelines. The pattern parallels MCP (agent-readable tool descriptions) but at the Spaces layer, which is where a large fraction of open-weight model demos already live. The meaningful question: does this stay a HuggingFace-native convention or does it generalize? If it does, the integration barrier that currently limits multimedia agent pipelines drops significantly.