excalidraw

Generating individual elements such as rectangles, text labels, and basic line parameters works adequately. The agent successfully outputs standard shapes and varied line styles, producing file structures that render without parsing crashes.

However, defining arrow-to-shape bindings works poorly. The agent must calculate relative visual offsets and generate matching identifier strings inside the JSON manually, which is highly sensitive to minor coordinate errors. Because the environment does not provide visual preview capabilities, the agent operates blindly and cannot check if its connectors align with target shapes.

Structuring complex container frames or grouped sections works poorly. Because there is no local schema verification, the agent outputs incorrect container properties, resulting in zero recognized grouping frames in the rendered artifact.

Iterative editing across separate sessions works poorly. When a fresh agent resumes work from a saved JSON canvas, the lack of structural metadata or inline schema guidance makes the raw dimensions difficult to interpret, often causing the agent to regress prior work when applying updates.

This tool is an adequate fit for agents generating simple, flat flowcharts or block layouts where only basic shapes and annotations are needed in a single turn. It is a poor fit for production-grade diagrams demanding custom container frames, exact arrow alignments, or collaborative, multi-step design updates.

The following diagram illustrates the open-loop generation process where the lack of an interactive rendering loop or schema linter leads to undetected structural errors: