Comment by avereveard

The differentiating thing that llm tool calls can't do reliably is to handle a lot of data. if tool a emit data that tool b needs, and it's a significant compared to model context, scripting these tool to be chained in a code fragment where they are exposed as functions saves a lot of pain