Software developers increasingly rely on Large Language Models (LLMs) to generate code that uses external libraries to accelerate development and avoid reinventing the wheel. However, LLMs struggle with the correct use of external libraries, particularly during the import of libraries, and frequently introduce errors that lead to time-consuming debugging and reduced developer productivity.

This work focuses on reducing library-related errors in LLM-generated code. Our contributions are twofold: first, we conduct an exploratory study to identify and quantify the main error patterns LLMs produce during library imports; second, we propose an agentic approach to address these issues. In a preliminary evaluation with 50 Python code generation tasks derived from real-world scripts and using GPT-5.2, our agentic system nearly halved library–related errors, demonstrating its potential to improve LLM-generated code reliability.