LLM as General purpose compute

Things that used to be orchestration on enterprise service buses are now Multi agent Systems

Also being used as classifiers but within the larger context of orchestrations that essentially makes them a big if statement

UPDATE 2026-Oct-1: Jev is turning this into a first-class construct--LLM-as-classifier.

When you look at the prompts they can sometimes look like very poorly specified programs. But they forget that LLMs aren't symbolic, nor deterministic. If you give it conditions under which it should return a tool call, it might get it right 95% of the time, but you might be better off chain that into a much shorter response classify prompt or just some simple sub string matching connected to a real statement that calls the tools.

don't try to do it all in one prompt

On the other hand, when is good enough good enough?