Something about agent.python_executor and additional_authorized_imports

[baixianger]

I have a question about a code snippet from the smolagents web browser automation example:

from smolagents import CodeAgent
from smolagents.agents import ActionStep

# Create the agent
agent = CodeAgent(
    tools=[go_back, close_popups, search_item_ctrl_f],
    model=model,
    additional_authorized_imports=["helium"],
    step_callbacks=[save_screenshot],
    max_steps=20,
    verbosity_level=2,
)

# Import helium for the agent
agent.python_executor("from helium import *", agent.state)

In this code, there are two instances where the helium library is referenced:
1. The additional_authorized_imports=["helium"] parameter in the CodeAgent initialization.
2. The agent.python_executor("from helium import *", agent.state) command.

I understand that additional_authorized_imports is used to grant the agent permission to use specific packages in the code it generates. Does this mean that if a tool within the agent requires the helium library, it doesn’t need to be listed here, and this parameter is solely for imports in the agent’s own code?

Additionally, could someone clarify the purpose of the second reference, agent.python_executor("from helium import *", agent.state)? Is this command necessary to explicitly import the helium functions into the agent’s execution environment, allowing it to utilize them during its operations?

Thank you for your assistance in understanding these distinctions.

ref: https://huggingface.co/docs/smolagents/en/examples/web_browser
Git

[yashwant938]

Yes, for the default local executor, those two lines serve different purposes:

1. additional_authorized_imports=["helium"] allows imports. It permits import helium / from helium import ... inside code evaluated by the executor, whether the model generated that code or you passed it to agent.python_executor(...). It does not install Helium or bind its functions in the executor's namespace.
2. agent.python_executor("from helium import *") performs the import. It puts the exported names into the executor's persistent state, so subsequent code can use names such as go_to directly. Importing Helium in your surrounding Python script alone does not populate that state.

For your tool question: a registered local Python tool executes its implementation as a normal Python callable. Its internal imports do not need to be added to the executor's import list solely because the tool uses them. Its dependencies still need to be installed in the host environment. This distinction is specifically about local tools; remote execution has its own environment requirements.

There is also a version detail in the example: with smolagents 1.26.0, the call takes just the code string:

# After constructing the agent with additional_authorized_imports=["helium"]:
agent.python_executor("from helium import *")

The older agent.python_executor("from helium import *", agent.state) form raises TypeError on that version. The current web-agent implementation already uses the one-argument form; the linked documentation example still shows the older call.

Preloading is optional if the agent's generated code imports what it needs itself. In this particular example, the instructions tell the model that the import has already happened, so keeping that preload makes the instructions match the execution state. The executor's import handling shows the authorization check and the separate state assignments.

Verification: six local checks passed on both the PyPI 1.26.0 package and a 1.27.0.dev0 checkout based on 96f33fa (Ubuntu/WSL, Python 3.14.4). They cover permission without import, rejected unauthorized imports, preloading across agent runs, imports in generated code, imports inside a registered tool, and the old call signature. The checks used the standard-library fractions module and a scripted model; they did not run Helium, a browser, or a model-provider API. Prepared with AI assistance and checked against the source and those test runs.