smolagents/docs/source/guided_tour.md

Agents - Guided tour

open-in-colab

In this guided visit, you will learn how to build an agent, how to run it, and how to customize it to make it work better for your use-case.

Building your agent

To initialize a minimal agent, you need at least these two arguments:

• An LLM to power your agent - because the agent is different from a simple LLM, it is a system that uses a LLM as its engine.
• A list of tools from which the agent pick tools to execute

For defining your LLM, you can make a custom_model method which accepts a list of messages and returns text. This callable also needs to accept a stop_sequences argument that indicates when to stop generating.

from huggingface_hub import login, InferenceClient

login("<YOUR_HUGGINGFACEHUB_API_TOKEN>")

model_id = "Qwen/Qwen2.5-72B-Instruct"

client = InferenceClient(model=model_id)

def custom_model(messages, stop_sequences=["Task"]) -> str:
    response = client.chat_completion(messages, stop=stop_sequences, max_tokens=1000)
    answer = response.choices[0].message.content
    return answer

You could use any custom_model method as long as:

1. it follows the messages format (List[Dict[str, str]]) for its input messages, and it returns a str.
2. it stops generating outputs at the sequences passed in the argument stop_sequences

Additionally, custom_model can also take a grammar argument. In the case where you specify a grammar upon agent initialization, this argument will be passed to the calls to model, with the grammar that you defined upon initialization, to allow constrained generation in order to force properly-formatted agent outputs.

For convenience, we provide pre-built classes for your model engine:

• [TransformersModel] takes a pre-initialized transformers pipeline to run inference on your local machine using transformers.
• [HfApiModel] leverages a huggingface_hub.InferenceClient under the hood.
• We also provide [LiteLLMModel], which lets you call 100+ different models through LiteLLM!

You will also need a tools argument which accepts a list of Tools - it can be an empty list. You can also add the default toolbox on top of your tools list by defining the optional argument add_base_tools=True.

Once you have these two arguments, tools and model, you can create an agent and run it.

from smolagents import CodeAgent, HfApiModel

model = HfApiModel(model_id=model_id)
agent = CodeAgent(tools=[], model=model, add_base_tools=True)

agent.run(
    "Could you give me the 118th number in the Fibonacci sequence?",
)

Note that we used an additional additional_detail argument: you can additional kwargs to agent.run(), they will be baked into the prompt as text.

You can use this to indicate the path to local or remote files for the model to use:

agent = CodeAgent(tools=[], model=model, add_base_tools=True)

agent.run("Why does Mike not know many people in New York?", audio="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/recording.mp3")

It's important to explain as clearly as possible the task you want to perform. Since an agent is powered by an LLM, minor variations in your task formulation might yield completely different results. You can also run an agent consecutively for different tasks: if you leave the default option of True for the flag reset when calling agent.run(task), the agent's memory will be erased before starting the new task.

Code execution

A Python interpreter executes the code on a set of inputs passed along with your tools. This should be safe because the only functions that can be called are the tools you provided (especially if it's only tools by Hugging Face) and a set of predefined safe functions like print or functions from the math module, so you're already limited in what can be executed.

The Python interpreter also doesn't allow imports by default outside of a safe list, so all the most obvious attacks shouldn't be an issue. You can authorize additional imports by passing the authorized modules as a list of strings in argument additional_authorized_imports upon initialization of your [CodeAgent] or [CodeAgent]:

from smolagents import CodeAgent

agent = CodeAgent(tools=[], model=model, additional_authorized_imports=['requests', 'bs4'])
agent.run("Could you get me the title of the page at url 'https://huggingface.co/blog'?")

This gives you at the end of the agent run:

'Hugging Face – Blog'

The execution will stop at any code trying to perform an illegal operation or if there is a regular Python error with the code generated by the agent. You can also use E2B code executor instead of a local Python interpreter by passing use_e2b_executor=True upon agent initialization.

[!WARNING] The LLM can generate arbitrary code that will then be executed: do not add any unsafe imports!

The system prompt

Upon initialization of the agent system, a system prompt (attribute system_prompt) is built automatically by turning the description extracted from the tools into a predefined system prompt template.

But you can customize it!

Let's see how it works. For example, check the system prompt for the [CodeAgent] (below version is slightly simplified).

The prompt and output parser were automatically defined, but you can easily inspect them by calling the system_prompt_template on your agent.

print(agent.system_prompt_template)

Here is what you get:

You are an expert assistant who can solve any task using code blobs. You will be given a task to solve as best you can.
To do so, you have been given access to a list of tools: these tools are basically Python functions which you can call with code.
To solve the task, you must plan forward to proceed in a series of steps, in a cycle of 'Thought:', 'Code:', and 'Observation:' sequences.

At each step, in the 'Thought:' sequence, you should first explain your reasoning towards solving the task and the tools that you want to use.
Then in the 'Code:' sequence, you should write the code in simple Python. The code sequence must end with '<end_code>' sequence.
During each intermediate step, you can use 'print()' to save whatever important information you will then need.
These print outputs will then appear in the 'Observation:' field, which will be available as input for the next step.
In the end you have to return a final answer using the `final_answer` tool.

Here are a few examples using notional tools:
---
{examples}

Above example were using notional tools that might not exist for you. On top of performing computations in the Python code snippets that you create, you only have access to these tools:

{{tool_descriptions}}

{{managed_agents_descriptions}}

Here are the rules you should always follow to solve your task:
1. Always provide a 'Thought:' sequence, and a 'Code:\n```py' sequence ending with '```<end_code>' sequence, else you will fail.
2. Use only variables that you have defined!
3. Always use the right arguments for the tools. DO NOT pass the arguments as a dict as in 'answer = wiki({'query': "What is the place where James Bond lives?"})', but use the arguments directly as in 'answer = wiki(query="What is the place where James Bond lives?")'.
4. Take care to not chain too many sequential tool calls in the same code block, especially when the output format is unpredictable. For instance, a call to search has an unpredictable return format, so do not have another tool call that depends on its output in the same block: rather output results with print() to use them in the next block.
5. Call a tool only when needed, and never re-do a tool call that you previously did with the exact same parameters.
6. Don't name any new variable with the same name as a tool: for instance don't name a variable 'final_answer'.
7. Never create any notional variables in our code, as having these in your logs might derail you from the true variables.
8. You can use imports in your code, but only from the following list of modules: {{authorized_imports}}
9. The state persists between code executions: so if in one step you've created variables or imported modules, these will all persist.
10. Don't give up! You're in charge of solving the task, not providing directions to solve it.

Now Begin! If you solve the task correctly, you will receive a reward of $1,000,000.

The system prompt includes:

• An introduction that explains how the agent should behave and what tools are.
• A description of all the tools that is defined by a {{tool_descriptions}} token that is dynamically replaced at runtime with the tools defined/chosen by the user.
  • The tool description comes from the tool attributes, name, description, inputs and output_type, and a simple jinja2 template that you can refine.
• The expected output format.

You could improve the system prompt, for example, by adding an explanation of the output format.

For maximum flexibility, you can overwrite the whole system prompt template by passing your custom prompt as an argument to the system_prompt parameter.

from smolagents import ToolCallingAgent, PythonInterpreterTool, TOOL_CALLING_SYSTEM_PROMPT

modified_prompt = TOOL_CALLING_SYSTEM_PROMPT

agent = ToolCallingAgent(tools=[PythonInterpreterTool()], model=model, system_prompt=modified_prompt)

[!WARNING] Please make sure to define the {{tool_descriptions}} string somewhere in the template so the agent is aware of the available tools.

Inspecting an agent run

Here are a few useful attributes to inspect what happened after a run:

• agent.logs stores the fine-grained logs of the agent. At every step of the agent's run, everything gets stored in a dictionary that then is appended to agent.logs.
• Running agent.write_inner_memory_from_logs() creates an inner memory of the agent's logs for the LLM to view, as a list of chat messages. This method goes over each step of the log and only stores what it's interested in as a message: for instance, it will save the system prompt and task in separate messages, then for each step it will store the LLM output as a message, and the tool call output as another message. Use this if you want a higher-level view of what has happened - but not every log will be transcripted by this method.

Tools

A tool is an atomic function to be used by an agent.

You can for instance check the [PythonInterpreterTool]: it has a name, a description, input descriptions, an output type, and a __call__ method to perform the action.

When the agent is initialized, the tool attributes are used to generate a tool description which is baked into the agent's system prompt. This lets the agent know which tools it can use and why.

Default toolbox

Transformers comes with a default toolbox for empowering agents, that you can add to your agent upon initialization with argument add_base_tools = True:

• DuckDuckGo web search*: performs a web search using DuckDuckGo browser.
• Python code interpreter: runs your the LLM generated Python code in a secure environment. This tool will only be added to [ToolCallingAgent] if you initialize it with add_base_tools=True, since code-based agent can already natively execute Python code
• Transcriber: a speech-to-text pipeline built on Whisper-Turbo that transcribes an audio to text.

You can manually use a tool by calling the [load_tool] function and a task to perform.

from transformers import load_tool

search_tool = load_tool("web_search")
print(search_tool("Who's the current president of Russia?"))

Create a new tool

You can create your own tool for use cases not covered by the default tools from Hugging Face. For example, let's create a tool that returns the most downloaded model for a given task from the Hub.

You'll start with the code below.

from huggingface_hub import list_models

task = "text-classification"

model = next(iter(list_models(filter=task, sort="downloads", direction=-1)))
print(model.id)

This code can quickly be converted into a tool, just by wrapping it in a function and adding the tool decorator:

from transformers import tool

@tool
def model_download_tool(task: str) -> str:
    """
    This is a tool that returns the most downloaded model of a given task on the Hugging Face Hub.
    It returns the name of the checkpoint.

    Args:
        task: The task for which
    """
    model = next(iter(list_models(filter="text-classification", sort="downloads", direction=-1)))
    return model.id

The function needs:

• A clear name. The name usually describes what the tool does. Since the code returns the model with the most downloads for a task, let's put model_download_tool.
• Type hints on both inputs and output
• A description, that includes an 'Args:' part where each argument is described (without a type indication this time, it will be pulled from the type hint). All these will be automatically baked into the agent's system prompt upon initialization: so strive to make them as clear as possible!

[!TIP] This definition format is the same as tool schemas used in apply_chat_template, the only difference is the added tool decorator: read more on our tool use API here.

Then you can directly initialize your agent:

from smolagents import CodeAgent
agent = CodeAgent(tools=[model_download_tool], model=model)
agent.run(
    "Can you give me the name of the model that has the most downloads in the 'text-to-video' task on the Hugging Face Hub?"
)

You get the following:

======== New task ========
Can you give me the name of the model that has the most downloads in the 'text-to-video' task on the Hugging Face Hub?
==== Agent is executing the code below:
most_downloaded_model = model_download_tool(task="text-to-video")
print(f"The most downloaded model for the 'text-to-video' task is {most_downloaded_model}.")
====

And the output: "The most downloaded model for the 'text-to-video' task is ByteDance/AnimateDiff-Lightning."

Multi-agents

Multi-agent has been introduced in Microsoft's framework Autogen. In this type of framework, you have several agents working together to solve your task instead of only one. It empirically yields better performance on most benchmarks. The reason for this better performance is conceptually simple: for many tasks, rather than using a do-it-all system, you would prefer to specialize units on sub-tasks. Here, having agents with separate tool sets and memories allows to achieve efficient specialization. For instance, why fill the memory of the code generating agent with all the content of webpages visited by the web search agent? It's better to keep them separate.

You can easily build hierarchical multi-agent systems with smolagents.

To do so, encapsulate the agent in a [ManagedAgent] object. This object needs arguments agent, name, and a description, which will then be embedded in the manager agent's system prompt to let it know how to call this managed agent, as we also do for tools.

Here's an example of making an agent that managed a specific web search agent using our [DuckDuckGoSearchTool]:

from smolagents import CodeAgent, HfApiModel, DuckDuckGoSearchTool, ManagedAgent

model = HfApiModel()

web_agent = CodeAgent(tools=[DuckDuckGoSearchTool()], model=model)

managed_web_agent = ManagedAgent(
    agent=web_agent,
    name="web_search",
    description="Runs web searches for you. Give it your query as an argument."
)

manager_agent = CodeAgent(
    tools=[], model=model, managed_agents=[managed_web_agent]
)

manager_agent.run("Who is the CEO of Hugging Face?")

[!TIP] For an in-depth example of an efficient multi-agent implementation, see how we pushed our multi-agent system to the top of the GAIA leaderboard.

Talk with your agent and visualize its thoughts in a cool Gradio interface

You can use GradioUI to interactively submit tasks to your agent and observe its thought and execution process, here is an example:

from smolagents import (
    load_tool,
    CodeAgent,
    HfApiModel,
    GradioUI
)

# Import tool from Hub
image_generation_tool = load_tool("m-ric/text-to-image")

model = HfApiModel(model_id)

# Initialize the agent with the image generation tool
agent = CodeAgent(tools=[image_generation_tool], model=model)

GradioUI(agent).launch()

Under the hood, when the user types a new answer, the agent is launched with agent.run(user_request, reset=False). The reset=False flag means the agent's memory is not flushed before launching this new task, which lets the conversation go on.

Next steps

For more in-depth usage, you will then want to check out our tutorials:

• the explanation of how our code agents work
• this guide on how to build good agents.
• the in-depth guide for tool usage.