Existing chart generation techniques face the challenge of balancing efficiency and quality: traditional manual templating methods are not scalable enough, while generalized large-language model generation schemes often lack domain constraints resulting in poor code executability and low quality charts. Meanwhile, the existing datasets mostly focus on the chart quiz task and neglect the standardization of the generation process, which is difficult to support high-quality automated chart generation research. To this end, this paper proposes a three-stage fully automated chart generation framework ChartGen-Agent in combination with a multimodal large language model, which achieves end-to-end generation from data/topic to high-quality charts through semantics-driven template construction, constraint-enhanced generation mechanism, and multi-dimensional quality optimization closed-loop. The framework innovatively integrates domain expert knowledge and multi-dimensional evaluation system (layout, color, semantic adaptation, etc.) to build a closed-loop of “Generate - Evaluate - Optimize”, which significantly improves chart standardization and code executability. The results of the experiments show that the framework has a significant advantage in terms of parameter efficiency and can still generate high-quality charts even when the model size is reduced to 7B. The Data2Charts dataset built on the framework contains 10,000 data, codes, chart triples covering 400 topics in 8 major domains, and its strict quality control process ensures that the charts meet academic visualization standards. This study provides a systematic solution to balance efficiency and quality in the field of intelligent visualization, and at the same time provides a standardized dataset and scalable framework for cross-domain high-quality chart generation research. Our source code and prompts are available on https://github.com/love-jr/chartgen-agent .