multimodal models with environmental grounding may eventually have an analog [of affect]. text-only agents can’t.

I'd be curious to hear an explanation about why text can't. Or equivalently, why multimodal could.