We actually started off using the AutoGPT framework. There are a ton of remnants of that (tasks, steps) but we found the framework extremely limiting as we wanted to expand and do more complex things
For example, we're currently using a multi agent architecture where we have micro agents run to analyze SVGs, fill out dynamic autocompletes. This would have been really hard.
Frameworks like langchain are good for early prototyping, but it's too restricting when you want to push the limits