Andy Halliday

In my case, Brian, it was right.

I mean, the model was right.

My thinking was brilliant.

So this brings to mind that there are ways to architect an AI assistant that has multiple agents in it.

And I can visualize this in an N8N workflow where

the output of one agent is actually sent on to an evaluator or a critic agent.

And that critic agent then adds that perspective, the sort of the counterpoint to it, to the next, you know, combines the output of agent one and the critic agent.

And so you get this more balanced analysis.

But typically, when you're working with a chatbot, you know, you don't have that unless you

In line, you ask it to now critique your own response and give me the negative view of that.

So I'm asking you now specifically, have you ever built a model, whether it's a chat GPT that calls another chat GPT, a custom GPT in this case, calls another one that's doing evaluative or critical analysis of what the original one was?

And what's the best way to build that?

So I want to bring- I wanted to say that I believe that you can, by building a very more complex structured prompt, you can drive the output of a single agent, if you will, or a single model's response to include, you know,

stages of evaluation and refinement of its thinking.

But we're moving towards a different model architecturally, I think, which is a mixture of agents where you create an agent whose specific skillset is for this kind of critique of the output of others and evaluation and testing of premises and so on.

But I just wanted to point out that in GenSpark,

they have a mode that you can select called mixture of agents but what it does is it just it just sends the same prompt the same query that you have to three different models right and then it gets the the reasoning output of each of those three different models and then it synthesizes those but it has a tendency to you know decide that one of the models is right

And then it filters the other models output, but because it's only giving the same prompt and it's not driving the other models to do something that's differential from any one.

So it's just, okay, here's three different outputs and we're going to kind of blend them together in order to do a,

a synthesis of that information and then give you the response.