Which is perhaps why the author was focussing on optimising LLM usage of such unwieldy APIs without blowing out the context ?

I don’t disagree with your point about agent abilities with older, concise, well-represented-in-training-data tools though.