What is agents.md? How is it different from OpenAPI or an SDK?

agents.md is a plain-text endpoint that Hugging Face provides for every Gradio Space — a single Markdown document that lays out the API schema query path, POST call format, how to poll results via event_id, file upload specs, and Bearer Token authentication. An agent only needs one GET request to read the whole thing and start calling it on its own, with no SDK to install and no integration logic to hardcode ahead of time. That makes it more directly consumable by an LLM than an OpenAPI JSON spec.

How do you use an AI agent to chain two Hugging Face Spaces into a 3D model?

The pipeline runs in two stages: first, call ideogram-ai/ideogram4 with a text prompt to generate a specimen-style image on a black background. Then upload that image via multipart to VAST-AI/TripoSplat, which reconstructs it into a 3D Gaussian Splat .ply file. Both Spaces expose their specs through agents.md, so the agent reads them, automatically polls the event_id for results, and finally embeds the .ply file into a frontend page for an interactive display.

Do you need to pay or apply for a quota to chain Hugging Face Spaces together?

Free Spaces come with CPU or shared GPU quota, but you'll deal with long queues and possible cold starts mid-run. For heavier workflows, it's worth upgrading to ZeroGPU or Pro and passing a Bearer Token with your calls to get higher priority. If you're treating a Space as a production-grade API, duplicate it into your own account and attach a paid GPU — otherwise you're exposed if the original author takes the Space down or changes it, breaking your whole pipeline.

What are the most common mistakes when chaining Spaces together?

Three come up most often: not handling the async queue properly, where the agent treats the event_id itself as the result; mismatched file formats, like an upstream Space outputting webp when the downstream one only accepts png; and ignoring cold starts, where the agent gives up as a failure just because the first call takes more than 30 seconds. It's worth building in polling timeouts, format conversion, and automatic retry-on-failure at the agent layer as safeguards.

Building-block economics versus training your own model — which fits which situation?

If the task is a combination of existing capabilities — text-to-image then image-to-3D, or speech-to-text then summarization — chaining ready-made Spaces is fastest, cutting development time from weeks to hours. Training your own fine-tuned model only pays off when you're dealing with proprietary data, extreme latency requirements, or business-sensitive IP. The upside of building-block economics is integration speed; the downside is you're dependent on upstream Spaces, so you need a fallback ready.

Who is this approach a good fit for?

Three groups in particular: independent developers who want to spin up demos fast, AI product teams that need multimodal pipelines, and consultants building hackathon or client POCs. A frontend engineer who knows how to write fetch calls and handle polling can treat thousands of Hugging Face Spaces as a backend to assemble from, without ever touching model training or GPU deployment. It's not a fit for real-time services with millisecond latency requirements.

Are there legal or copyright risks to using Spaces in a production product?

Spaces are uploaded by community authors, and licensing ranges from MIT to research-only, so you need to check the LICENSE and the underlying model's terms before any commercial use. Models like Ideogram and TripoSR each have their own rules around commercial output, and whether the generated images or 3D files can be sublicensed to a client depends on those terms. It's worth keeping a record of the original Space ID and model version in your pipeline for compliance tracing and easy source-swapping later.

AI Agent Chains Two Hugging Face Spaces to Auto-Generate a 3D Paris Gallery

This article is a deep-dive from JudyAI Lab — an AI engineering playbook series with 100+ published guides, 5,000+ weekly readers across 60+ countries, focused on the practical side of running AI agents, trading systems, and content pipelines in production.

📰 Key Takeaways

A developer had an AI agent independently produce every asset for a 3D showcase site of Paris landmarks — without manually opening a single image generation tool or 3D reconstruction program. The agent got there by directly chaining two Hugging Face Spaces: first calling ideogram-ai/ideogram4 to turn each landmark into a clean, specimen-style image on a black background using text prompts, then feeding that image into VAST-AI/TripoSplat, which reconstructed a 3D Gaussian Splat .ply file from the single image. The result was assembled into an interactive, cinematic showcase page.

The key piece of tech behind this is a plain-text agents.md endpoint that Hugging Face added to every Gradio Space. An agent just sends a GET request to fetch the document and gets the full calling spec back: the API schema query path, the POST call endpoint format, how to poll results via event_id, the multipart format for file uploads, and a Bearer Token authentication hint. No SDK required, no integration logic hardcoded in advance — the agent reads the doc and can drive the whole Space end to end.

The real breakthrough is the chaining itself: one Space’s output feeds directly into the next Space’s input, forming a complete “prompt → image → 3D model” pipeline. The author cites Mitchell Hashimoto’s idea of a “building-block economy” — AI isn’t great at building from scratch, but it’s excellent at snapping together components that already work. Hugging Face Spaces plus agents.md is exactly the infrastructure that turns multimedia AI models into blocks an agent can assemble directly.

💬 JudyAI Lab’s Take

When an agent can read the spec itself, chain the tools on its own, and deliver end to end, “AI assistant” quietly upgrades to “AI operator” — this case just crossed that line.

The key piece is the plain-text agents.md endpoint Hugging Face added to every Gradio Space: an agent needs just one GET request to pull the full API spec, no SDK required, no hardcoded integration logic upfront. What’s worth paying even closer attention to is the chaining design — one Space’s output feeds straight into the next Space’s input, forming a complete “prompt → image → 3D model” pipeline. The “building-block economy” concept the author cites points to a real direction: AI isn’t good at building from zero, but it’s excellent at assembling components that are already proven. For AI builders, the center of gravity in system design is shifting from “writing integration code” to “making every tool expose a spec an agent can read directly.”

Go check whether the AI tools you use regularly have a machine-readable spec endpoint, and find two existing tools whose output and input formats can chain together directly.

📅 Source Info

Published: 2026-06-09T10:46
Original source: https://huggingface.co/blog/mishig/spaces-agents-md

AI Agent Chains Two Hugging Face Spaces to Auto-Generate a 3D Paris Gallery

📰 Key Takeaways

💬 JudyAI Lab’s Take

📅 Source Info

🔗 Further Reading

References

📰 Key Takeaways#

💬 JudyAI Lab’s Take#

📅 Source Info#

🔗 Further Reading#

References#

Get our weekly AI digest:

📰 Key Takeaways

💬 JudyAI Lab’s Take

📅 Source Info

🔗 Further Reading

References