Huggingface AI Agents Quiz Solutions

I have been diving into AI agents through Huggingface’s AI Agents Course. This course offers a comprehensive understanding of how to build and deploy AI agents using the smolagents library. In this blog, I’ll share insights from the course (Unit 2) and provide code snippets to illustrate key concepts.

Note

Here is the course link if anyone is interested. AI Agents Course

Create a Basic Code Agent with Web Search Capability

One of the foundational exercises involves creating a CodeAgent equipped with web search capabilities. This agent leverages the DuckDuckGoSearchTool to perform web searches, enabling it to fetch real-time information. Here’s how you can set it up:

    # Create a CodeAgent with DuckDuckGo search capability
    from smolagents import CodeAgent, DuckDuckGoSearchTool, HfApiModel

    agent = CodeAgent(
        tools=[DuckDuckGoSearchTool()],           # Add search tool here
        model=HfApiModel("Qwen/Qwen2.5-Coder-32B-Instruct")          # Add model here
    )

In this snippet, we initialize a CodeAgent with the DuckDuckGoSearchTool, allowing the agent to perform web searches to answer queries.

Set Up a Multi-Agent System with Manager and Web Search Agents

Multi-Agent systems are the agents that are specialized on complex tasks with more scalable and robust nature. In smolagents, various agents can be integrated to produce Python code, invoke external tools, conduct web searches, and more. By coordinating these agents, it’s possible to develop robust workflows. A typical multi-agent system includes:

    - A manager Agent
    - A code interpreter Agent
    - A web Search Agent

Multi-agent system allows to separate memories between different sub-tasks and provide great benefits. Firstly, each agent are more focused on its core taks and secondly, separating memories reduces the count of input tokens resulting in reducing latency and cost. Below is the multi-agent system when web_agent performs search and manager_agent gives data analysis capabilities. Also, we can import dependencies (like python libraries) that helps to perform the tasks.

    from smolagents import CodeAgent, ToolCallingAgent, DuckDuckGoSearchTool, HfApiModel, VisitWebpageTool

    web_agent = ToolCallingAgent(
        tools=[DuckDuckGoSearchTool(), VisitWebpageTool()], 
        model=HfApiModel(model_id="Qwen/Qwen2.5-Coder-32B-Instruct"),
        max_steps=10,
        name="search", 
        description="Agent to perform web searches and visit webpages."
    )


    manager_agent = CodeAgent(
        model=HfApiModel(model_id="Qwen/Qwen2.5-Coder-32B-Instruct"),
        managed_agents=[web_agent],
        additional_authorized_imports=["pandas", "time", "numpy"] # Corrected imports
    )

Configure Agent Security Settings

Security is a crucial aspect when deploying AI agents, especially when they execute code. Below code snippet uses E2B to run code in a sandboxed environment. It is a remote execution that run the code in a isolated container.

    from smolagents import CodeAgent, HfApiModel
    from smolagents.sandbox import E2BSandbox

    model = HfApiModel("Qwen/Qwen2.5-Coder-32B-Instruct")

    agent = CodeAgent(
        tools=[],
        model=model,
        sandbox=E2BSandbox(),  # Configure the sandbox
        additional_authorized_imports=["numpy"],  # Authorize numpy import
    )

Implement a Tool-Calling Agent

Similar to CodeAgent, ToolCallingAgent is another type of agent available in smolagent library. CodeAgent uses Python code snippets whereas ToolCallingAgent use built-in tool-calling capabilities of LLM providers and generate JSON structures.

    from smolagents import ToolCallingAgent, HfApiModel, DuckDuckGoSearchTool

    agent = ToolCallingAgent(
        tools=[DuckDuckGoSearchTool()],  
        model=HfApiModel(model_id="Qwen/Qwen2.5-Coder-32B-Instruct"),  
        name="SearchAgent",  
        description="An agent that uses DuckDuckGo to search the web.",  
        max_steps=5,  
    )

Set Up Model Integration

LLM models are the most important aspect when creating AI agents. There are many model availables for various tasks and domains. So we can easily integrate models that is required for our task. Below code snippet switches between two different models providers.

    from smolagents import HfApiModel, LiteLLMModel

    # Initialize Hugging Face model
    hf_model = HfApiModel(model_id="Qwen/Qwen2.5-Coder-32B-Instruct")

    # Initialize LiteLLM model as an alternative model
    other_model = LiteLLMModel(model_id="anthropic/claude-3-sonnet")

    # Set the model to hf_model or alternative model
    model = hf_model  # Alternatively, you can switch this to `other_model`