Agentic AI systems increasingly perform data analysis workflows that require generating executable visualizations.
In primitive-based plotting libraries, agents must infer the semantic intent of a visualization and also synthesize the procedural code needed to construct figures, axes, legends, layouts, and interactions.
This paper studies whether a declarative visualization grammar reduces this interface-induced procedural burden.
We compare Matplotlib, as a primitive-based plotting interface, with Altair/Vega-Lite, as a declarative visualization interface.
Across 11 visualization tasks, two library conditions, three repeated runs, and four successfully evaluated OpenAI models, we measure first-pass generation tokens, generated code lines, execution success, repair-token behavior, and repeat-run variability.
The results show no consistent Altair advantage for low-complexity charts. However, for medium and higher-complexity tasks, Altair reduces first-pass generation tokens in most conditions, with aggregate reductions of 22.6\% for medium tasks, 12.1\% for medium-high tasks, and 31.1\% for high-complexity tasks.
Code-line reductions follow the same pattern. These results suggest that token efficiency is not merely a syntactic property of a library.
It emerges from the alignment between task semantics, tool semantics, and model capability.