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.