I was also under the impression modern AI agents have moved on from just OCR'ing screenshots to leveraging native vision model capabilities.