Data visualization generation from natural language queries (NL2VIS) has emerged as a critical research area, enabling non-technical users to create insights from structured data. While recent approaches have leveraged Large Language Models (LLMs) for end-to-end NL2VIS tasks, they suffer from the “lost in the middle” problem, generating incorrect visualization grammar for complex queries and requiring extensive post processing of unstructured outputs. This study introduces a hybrid agent framework that decomposes NL2VIS into structured subtasks, combining LLM understanding with formal grammar constraints and schema validation. Experimental evaluation on the VisEval benchmark dataset demonstrates significant performance improvements, with our agent-based system achieving 50.64% exact match accuracy and 99.74% execution success rate, surpassing existing LLM-based approaches by 18.6 …