Smolagents

Smolagents Skill

This skill leverages Hugging Face's Smolagents framework, a minimalist AI agent library where agents write Python code to accomplish tasks. It's highly efficient (30% token efficiency gain) and supports multiple LLM providers.

Setup

1. Dependencies: Requires smolagents with toolkit extensions.

   pip install 'smolagents[toolkit]' python-dotenv
   

2. API Key Configuration: Supports multiple LLM providers. At minimum, set one:

   # For Hugging Face Inference API (default, free tier available)
   echo "HF_TOKEN=your_huggingface_token" >> .env
   
   # OR for OpenAI
   echo "OPENAI_API_KEY=your_openai_key" >> .env
   
   # OR for Anthropic
   echo "ANTHROPIC_API_KEY=your_anthropic_key" >> .env
   
   if [ -f .gitignore ] && ! grep -q ".env" .gitignore; then echo ".env" >> .gitignore; fi
   

Usage

Use the scripts/agent.py script to run research tasks.

Command

python3 scripts/agent.py --task "<task_description>" [--model <model_type>] [--model-id <model_name>] [--web-search]

Parameters

• --task (Required): The task or research question.
• --model (Optional): Model type - hf (Hugging Face), openai, anthropic, or local (default: hf).
• --model-id (Optional): Specific model ID to use.
• --web-search (Optional): Enable web search tool (uses DuckDuckGo).
• --verbose (Optional): Show detailed execution logs.

Example

# Using Hugging Face Inference API with web search
python3 scripts/agent.py --task "Research the latest developments in transformer architecture improvements" --web-search --verbose

# Using a specific model
python3 scripts/agent.py --task "Analyze the impact of RLHF on LLM performance" --model hf --model-id "Qwen/Qwen2.5-72B-Instruct" --web-search

Output

The script outputs:

• Generated Python code (to stderr for visibility)
• Task execution results
• Final answer or research findings

Features

• Code-as-Action: Agents write and execute Python code to solve tasks
• Tool Support: Web search, file operations, and custom tools
• Multi-Model: Supports HF Inference API, OpenAI, Anthropic, local models
• Efficient: 30% token efficiency improvement over traditional approaches