Hi Google AI Team,
We are integrating Gemini models into our application and support two ways of accessing them:
-
Gemini Developer API — https://ai.google.dev/gemini-api/docs/code-execution
-
Gemini on Vertex AI / Agent Platform — Code execution | Gemini Enterprise Agent Platform | Google Cloud Documentation
We noticed that the built-in Python code execution environments have different sets of preinstalled libraries, according to the official documentation.
For example, the Gemini Developer API environment includes libraries such as python-pptx, python-docx, openpyxl, Pillow, and lxml. However, these libraries are not included in the documented list of supported packages for code execution on the Agent Platform, which provides a more limited set of libraries.
Our application uses Gemini models to perform tasks that involve generating documents and files, including PowerPoint presentations (PPTX). These workflows may require libraries such as python-pptx, which are available through the direct Gemini API’s code execution environment but are not listed as available in the Agent Platform environment.
Since we want to support both providers in our application, this difference in package availability creates a limitation when executing the same tasks through the Agent Platform.
We would appreciate clarification on the following questions:
-
Package availability: Why do the built-in code execution environments have different sets of preinstalled Python libraries between the Gemini Developer API and the Agent Platform?
-
Future support: Is there a roadmap or any plan to bring additional Python packages, such as
python-pptx,python-docx,openpyxl, andPillow, to the Agent Platform’s built-in code execution environment? -
Custom dependencies: Is there any supported way to install additional Python packages or configure the execution environment with custom dependencies?
-
Recommended approach: If expanding the built-in package set is not currently planned, what is Google’s recommended approach for applications that need to execute Python code using additional libraries while continuing to access Gemini models through the Agent Platform?
We understand that the built-in code execution environment is managed by Google and does not support installing arbitrary libraries. Our main question is whether broader package support is planned and whether there is a recommended way to handle these requirements in the meantime.
Any clarification regarding the roadmap or recommended architecture would be greatly appreciated.
Thank you!