You are an absolute moron for believing in the hype of "AI Agents"
nexustrade.io
nexustrade.io
Funny I worked on a "model trainer" circa 2018 which was a hot mess. The trouble with it was that some models could be trained in 30 seconds on a small machine and the tools from scikit-learn for model selection were great. Other models would take a few days and had huge data sets and we'd have to code up our own model selection tools. Over time I came to understand the model selection problem better and appreciate the tools in scikit-learn but it wasn't easy to fit those into a framework that involved working across many different address spaces over time.
It sucked and I blamed the complexity of the problem about 50% and my own ignorance about 50%. I've seen a lot of other flawed (but useful) tools since as the Huggingface tools (nice models, shame about the model selection) and other tools that are just crap (langchain... how they can hypnotize people to think there is any value there is beyond me) I am putting more blame on the complexity of the problem.
(Note we believed in foundation models back then, I did in 2009, and we wanted one system that would let us handle both those huge models and also rather simple models that build on them)
That model trainer worked with a tool that looked a lot like
but had transferred JSON data instead of rows over the lines (most tools won't do that because of speed & scalability concern so things that should be simple involve graphical joins that are far more mind melting than plain SQL joins) and also had AI-powered blocks trained by that trainer.
Another kinda graphical tool for AI agents would be something like
which would let you sketch out a workflow and have blocks that could be human or AI tasks. Specifically the human can fix it when the AI flubs it or thinks it can't do it, and the human can build up a training set for it's "digital twin". I had the damnedest time trying to sell those last two ideas to customers and investors, either could have been world changing, but instead people give non-answers like langchain so much more respectability than they deserve. (e.g. in 2001 I had an open-source user management system modeled after what Yahoo and AMZN had, nothing like that got popular in 2013 when I discovered that "you don't own your user list" and "locked into a vendor who will either go out of business or get bought by Google" where the critical features customers were demanding that I didn't have)