README.md

metadata

title: Alfred - HF Agents Course Unit 1
emoji: 🦇
colorFrom: indigo
colorTo: yellow
sdk: gradio
sdk_version: 5.49.1
app_file: app.py
python_version: '3.12'
pinned: false
license: apache-2.0
tags:
  - agent
  - smolagents
  - agents-course
  - unit-1

Unit 1 — First Agent ("Alfred")

A working smolagents CodeAgent with five tools:

Tool	Source
web_search	DuckDuckGoSearchTool() (smolagents default)
visit_webpage	VisitWebpageTool() (smolagents default)
get_current_time_in_timezone	@tool-decorated, custom (pytz)
convert_currency	@tool-decorated, custom (exchangerate.host)
final_answer	FinalAnswerTool() (terminal)

Run locally:

pip install -r requirements.txt
export HF_TOKEN=hf_...
python app.py

Deploying to a Space:

1. Create a new Gradio Space (or duplicate agents-course/First_agent_template).
2. Copy app.py + requirements.txt into the Space repo.
3. Settings → Variables and Secrets → New Secret: HF_TOKEN = your inference-enabled token.
4. Push. The Space will build, and its UI is launched by GradioUI(agent).launch().

Sample prompts that exercise each tool

• "What time is it right now in Tokyo and Paris?" → get_current_time_in_timezone
• "Convert 500 USD to EUR and to JPY." → convert_currency
• "What is smolagents and why does HuggingFace recommend it?" → web_search + visit_webpage
• "What's the time in New York and how many EUR is 1000 USD worth right now?" → mixed multi-step

Notes / lessons applied

• InferenceClientModel(model_id=...) defaults to Qwen 2.5-Coder-32B. Override with the AGENT_MODEL_ID env var if you want a different one.
• max_steps=6 is the safety hatch for the ReAct loop.
• The final_answer tool is kept in the tools list — without it the loop has no terminal action.
• The agent is shippable as-is to a Space (Gradio UI bundled via GradioUI(agent).launch()).