matplotlib

The agent designs standard visualizations using Matplotlib's object-oriented interface. Standard methods like subplots, bar charts, and scatter plots work adequately because their parameters correspond directly to data series and match predictable API patterns. Creating advanced multi-axis layouts using methods like twinx and applying styling options like hatch patterns also works adequately, as these properties can be configured explicitly in Python code.

However, constructing multi-element compositions like data cards and custom legends works poorly. The agent must position titles, borders, and text blocks by specifying explicit coordinate systems and bounding boxes. Because the agent lacks direct visual feedback, calculating precise spatial offsets is highly iterative and error-prone, where minor alterations frequently cause overlapping text or clipped borders without throwing structural errors.

State persistence across multiple sequential executions works poorly. Repeated attempts to run or modify code in an ongoing session frequently fail due to execution environment unreliability. Conversely, resuming development from a clean state works well, since a fresh agent resuming from saved code files can modify the layout without breaking previously established structures.

Matplotlib is a good fit for agents executing automated data visualization tasks that require standard charts and structured layouts defined through explicit code parameters. It is effective when the layout is simple or when exact coordinate alignment is not critical to the final output.

The tool is a poor fit for tasks that demand complex page designs, custom templates, or precise canvas layouts. When an agent must design visually dense media containing multi-axis charts, multi-column metric boxes, and precise text callouts, the lack of real-time visual feedback makes the process inefficient.

This diagram illustrates the agent's iterative write-execute-verify workflow when generating artifacts with Matplotlib.