designfailure/agents-md-agent

agents-md-agent — "Hub Scout"

An agent whose entire instruction set lives in one human-readable file, AGENTS.md. The code (app.py) is just the wiring: it loads AGENTS.md verbatim as the system prompt and gives the model two tools that hit the real Hugging Face Hub. Edit AGENTS.md and the agent's behavior changes — no code changes. This is the "author markdown, push, it's live" pattern.

Run locally

git clone https://huggingface.co/datasets/designfailure/agents-md-agent
cd agents-md-agent
pip install -r requirements.txt

export HF_TOKEN=hf_...        # token with Inference Providers access
python app.py                 # opens http://127.0.0.1:7860

Override the model with AGENT_MODEL=<repo id> (any tool-capable chat model available through Inference Providers).

Tools the agent can call

• —search_hf_models(query, limit, task) — keyword search over the Hub, ranked by likes
• —get_model_stats(repo_id) — downloads, likes, last-modified for one repo

Try asking it

• —"Find me a small chat model that runs on a laptop CPU."
• —"What's the most downloaded sentence-embedding model right now?"
• —"How actively used is Qwen/Qwen2.5-72B-Instruct?"

Portability

agent.json follows the deepagents project layout, so the same AGENTS.md drops straight into the deepagents harness or deepagents deploy (which needs a LangSmith key) without rewrites. The full file layout this expects:

agents-md-agent/
  agent.json        # name, description, model          (deepagents convention)
  AGENTS.md         # the agent's entire instruction set — the point of the project
  app.py            # Gradio wiring: system prompt = AGENTS.md, two real tools
  requirements.txt
  README.md