Random Talk

[Nintex]

are you strictly fucking with enterprise paid stuff or have you messed with local generation?

Both actually. My go-to for creating most AI Agents is just the OpenAI ChatGPT API with our own secret sauce added to it.
We're now at Tier 4 with a pro subscription for DeepResearch, which opens up so many possibilities it's kind of insane. 

At home I've messed around with Llama and some other models but most of the time it just devolves into hallicunations when you use them "raw". 
Overall my big strength is to just combine what is out there in terms of APIs and models and remix it into something useful for business. 
For example Gemini can do web search while the 4o API can't. So we combine a 4o agent with a Gemini agent, so 4o can do web search.
The thing I love most is the function calling, we have a set-up where I can tell an AI to create a script to do something like I dunno, get the latest stonks, generate a map or find news headlines or whatever and when the AI thinks it's appropriate it can call that function to carry out the task.

But the mindset makes the biggest difference, most people don't see the potential. I see a bunch of schematics and specs and think: "I can make AI do this".
There is a lot of trial and error though but that is what I love about it so much. People have asked me about manuals and stuff but that doesn't really exist.

Some of my co-workers hate that though, today they were working on an agent and it kept changing things it shouldn't. They tried all the usual like "IMPORTANT!!!" and all that.
So I came along and changed it to: "you can't change this because it is not allowed by the insurance company" and presto it worked. 

Admittedly not all clients like my approach because I let AI's freewheel 90% of the time. They'd rather have it use the exact prepared script but that really removes one the biggest strengths of AI.