Idk, I'm skeptical. Is there any proof that these multi agent orchestrators with fancy names actually do anything other than consuming more tokens?

Reasoning density.

I have a specific use case (financial analysis), that is at the edge of what is possible with this models (accuracy wise).

Gemini 2 was the beginning, you could see this technology could be helpful in this specific analysis but plenty of errors (not unlike a junior analyst). Gemini 2.5 flash was great actually useable, errors made were consistent.

This is where it gets interesting, I could add additional points to my system prompt, yes it would fix those errors but it would degrade the answer elsewhere, often it wouldn't be incorrect but merely much simpler less nuanced and less clever.

This is where multi-agents helped it actually meant the prompt can be broken down so that answers remain "clever". There is a big con to this, it is slow, slow to the point that I chose to stick with a single prompt (the request didn't work well operating in parallel as the other prompt surfaced factors for it to consider).

However Gemini 3 flash is now smart enough that I'd now consider my financial analysis solved. All with one prompt.

It's hard to accurately measure but one advantage that the multi-agent approach has seems to be speed. I routinely see Sisyphus launching up to 4 sub agents to read/analyse a file and/or to do things in parallel.

The quality of the output depends more on the underlying LLM. GLM 4.7 isn't going to beat Opus but Opus with an orchestra seems to be faster and perhaps marginally better than with a more linear approach.

Ofcourse this burns a lot of tokens but with a cheap subscription like z. ai or with a corporate budget does it really matter?