Yet Another "Tool Using" Therapeutic AI

Human-like or New Caledonian crow-like?

Tool use is what separates the animals – “zero-shot” LLMs – from the intellectually gifted – “agentic” AI – after all.

This is a Harvard/MIT demonstration project in which their “TxAgent” is designed to answer questions about medication use. In their agent, they’ve paired the LLM substrate with the ability to access 201 potential APIs across their “ToolUniverse”. The conceptual overview of the implementation is as follows:

The full article describes their manner of construction and their training process, but the gist is obvious – agents, properly designed, are better:

I don’t highlight these examples as products to pursue, but illustrative of the techniques and training approaches others might find fruitful, or for clinical leadership to review for greater understanding of the underlying structure.