Ask any frontier model to "draw an architecture diagram as SVG" and you'll get something that looks like markup and renders like a ransom note: boxes overlapping, labels spilling past their borders, arrows cutting straight through other shapes. The model wrote valid SVG.
It just can't see.
That's the whole problem in one sentence: a language model has no visual cortex. It predicts tokens, not pixels. Asking it to place a node at x=412, y=088 and route an edge around three other nodes is asking it to do collision detection and graph layout in its head,
blind, one token at a time. It will confidently get it wrong.
The two bad options agents have today
