Who decides at each fork in your workflow?

The oldest is the logic gate. If this, then that, otherwise something else. That is what we have seen in software for decades. If the mail comes from a fixed address, or if there is a specific word in the subject, such a rule is sufficient. It costs nothing and it never doubts.

Besides that, you now have the agentic decision gate. Then you let a language model decide, a Claude, a ChatGPT or a Gemini. That is handy when the input is different every time. A mail from a customer who is angry but never literally says so, will not be recognised by a rule. An agent will.

And then there is something new: models that are specially made to decide. Suppose you have a folder with many different types of documents, which all need to be read and neatly organised. You can have a classic language model do that, but that costs you a lot of tokens. A model that only decides does the same sorting work much faster and much more economically. For such a pile, that is actually the logical choice.

You might forget the fourth one, and that is you. The human is still in the story, and you can be a decision gate yourself. In my example, I decide myself whether the draft leaves. I would keep that role wherever something is irreversible, goes outside, or where you need a mandate.