excalidraw

The agent constructs the diagram by explicitly specifying coordinates, text bindings, and connection attributes in raw JSON. It defines properties like stroke, fill, and line bindings directly. This manual syntax creation becomes a verbose, expensive process that requires numerous tool calls to build even basic shapes.

A major friction point is the lack of a feedback loop for visual verification. Because the agent cannot render and view the resulting canvas, it generates layout parameters blindly. It cannot visually inspect whether text overflows shapes or if connectors overlap, making minor adjustment tasks highly error-prone.

Furthermore, working without local documentation forces the agent to rely on pre-existing knowledge of the JSON schema. While it successfully generates syntax-valid output that renders cleanly, it frequently misses semantic details. For example, it attempts to group regions using simple graphical rectangles rather than defining native frame objects, leaving schema-specific requirements unmet.

Finally, resuming the work with a fresh agent leads to regressions. When a new agent inherits the raw JSON file, it struggles to reconstruct the visual hierarchy and design intent without a rendering tool or schema template, causing it to damage or revert previously functioning components.

This format is suitable only for generating basic shapes and layout concepts that are simple enough to calculate programmatically. It works best when the agent receives a strict, pre-defined template schema to fill in rather than writing elements from scratch.

It is a poor fit for interactive diagrams, complex nested architectures, or tasks where visual alignment and semantic frame grouping are critical. Without live rendering utilities or schema-validation tools to inspect the layout, agents cannot guarantee visual correctness or satisfy complex layout rules.

The following diagram illustrates the agent's blind generation loop where it writes raw JSON coordinates from memory without any visual feedback or schema validation before producing the final asset.