Unless I've misunderstood, it is most effective on a picture of text and has to answer with text. It is extremely difficult for it to guide you through some GUI or give you a sequence you may want to correct a little without forcing you to study what exactly it is doing instead of cutting and pasting text into a text UI.
It's hard for me to imagine if multiple AGI wrapped interfaces could use some other input, I.e. emulated remote desktops and screen shares, (and that could be adequately chainable for AGI output to other interface input,) but I feel like adding all of this data is ultimately making it harder to proof read and adapt something AGI proposes and then automate its repeatable usage (like taking scripts or code.)
It's hard for me to imagine if multiple AGI wrapped interfaces could use some other input, I.e. emulated remote desktops and screen shares, (and that could be adequately chainable for AGI output to other interface input,) but I feel like adding all of this data is ultimately making it harder to proof read and adapt something AGI proposes and then automate its repeatable usage (like taking scripts or code.)