800x800 is 640,000 pixels, or 0.64 Megapixels. That is less than the resolution of computer screens from 1995, Super VGA which has around 0.79 MPs.
This is useful for a reasonable amount of use-cases, but I think the watershed rez will be around triple that, ~1080p, which is enough for almost anything, except small text and subtle details.
You'd expect a tool-enabled model to leverage crop and zoom tools to inspect and validate what it thinks it's seeing, though.
I typically provide small screenshots to llms so this seems fine for that usecase, providing an entire screens context seems cause confusion with a lot of llms.