Hi everyone,
I wanted to share a project I’ve been building: HF Agent OS, a small multi-agent platform that lets you search, inspect, run inference on, upload to, and deploy on the Hugging Face Hub — all through natural language.
Instead of one monolithic agent trying to do everything, it’s structured as:
smolagents.Tool
registry) shared across all agentsIt’s all built on smolagents, using CodeAgent
InferenceClientModel
for each sub-agent.
The part I’m most happy with: the Model Agent demonstrated genuine autonomous self-correction during testing. I asked it to run inference on a partial/invalid model ID (distilbert-base-uncased-finetuned-sst-2-english
, missing the distilbert/
namespace). Instead of just failing, it:
list_models
tool to search the Hub for the correct IDdistilbert/distilbert-base-uncased-finetuned-sst-2-english
No human intervention — it just recovered on its own.
On safety: any tool that can create or modify real Hub resources (repo creation, file upload, Space deployment) follows a strict preview-then-confirm pattern. A read-only preview tool always shows what would happen (target repo, file size, resulting URL) before the corresponding write tool is allowed to act, and the write tool only fires on an explicit, unambiguous user request.
Everything’s been tested against the real Hub — real repo creation, real file uploads, real inference calls — plus 74 passing unit tests with all Hub calls mocked.
One fun constraint I ran into: deploying a real Gradio Space requires a PRO subscription for compute-backed Spaces (Static Spaces are free) — good reminder that “deployment agent” work also means understanding the platform’s actual billing model, not just the API surface.
Code is on GitHub: GitHub - mebrouktalbi76-eng/hf-agent-os · GitHub Would love feedback, especially from anyone who’s built multi-agent systems with smolagents — curious how others have approached the planner/routing layer, and whether there’s appetite for something like this as a reusable pattern for other Hub-adjacent workflows.