Smolagents

​

Overview

​

Prerequisites

1. Before you can use Smolagents with Weave, you must install the required libraries or upgrade to the latest versions. The following command installs or upgrades smolagents, openai, and weave, and suppresses output:
   pip install -U smolagents openai weave -qqq
   
2. Smolagents supports multiple LLM providers, such as OpenAI, Hugging Face Transformers, and Anthropic. Set the API key for your chosen provider so that Smolagents can authenticate when invoking the model. Set the corresponding environment variable:
   import os
   import getpass
   
   os.environ["OPENAI_API_KEY"] = getpass.getpass("Enter your OpenAI API key: ")
   

​

Basic tracing

• You define a language model (OpenAI’s gpt-4o) using Smolagents’ OpenAIServerModel.
• You configure a search tool (DuckDuckGoSearchTool) that the agent can invoke when needed.
• You construct a ToolCallingAgent, passing in the tool and model.
• You run a query through the agent that triggers the search tool.
• Weave logs each function and model invocation, making them available for inspection in its web interface.

import weave
from smolagents import DuckDuckGoSearchTool, OpenAIServerModel, ToolCallingAgent

# Initialize Weave
weave.init(project_name="smolagents")

# Define your LLM provider supported by Smolagents
model = OpenAIServerModel(model_id="gpt-4o")

# Define a DuckDuckGo web search tool based on your query
search_tool = DuckDuckGoSearchTool()

# Define a tool-calling agent
agent = ToolCallingAgent(tools=[search_tool], model=model)
answer = agent.run(
    "Get me just the title of the page at url 'https://forge.coreweave.com/wandb/geekyrakshit/story-illustration/reports/Building-a-GenAI-assisted-automatic-story-illustrator--Vmlldzo5MTYxNTkw'?"
)

​

Tracing custom tools

• You define a custom get_weather function and decorate it with @tool from Smolagents, enabling the agent to invoke it as part of its reasoning process.
• The function accepts a location and an optional flag for Celsius output.
• You instantiate a language model using OpenAIServerModel.
• You create a ToolCallingAgent with the custom tool and model.
• When the agent runs the query, it selects and invokes the get_weather tool.
• Weave logs both the model inference and the custom tool invocation, including arguments and return values.

from typing import Optional

import weave
from smolagents import OpenAIServerModel, ToolCallingAgent, tool

weave.init(project_name="smolagents")

@tool
def get_weather(location: str, celsius: Optional[bool] = False) -> str:
    """
    Get the weather in the next few days for a given location.
    Args:
        location: The location.
        celsius: Whether to use Celsius for temperature.
    """
    return f"The weather in {location} is sunny with temperatures around 7°C."

model = OpenAIServerModel(model_id="gpt-4o")
agent = ToolCallingAgent(tools=[get_weather], model=model)
answer = agent.run("What is the weather in Tokyo?")