My first implementations were a mess with OpenClaw. I'll spare you the details. I moved to n8n. n8n was not bad but it felt like I had to strong arm it to do what I wanted. Had issues, but less.
Then this system - LangGraph / FastAPI... it has been working pretty well. The papers generated seem to have meaning. Generation of ideas have shown to have some positive substance. All ideas are grounded on a basis of pass / fail or positive / negative - based on criteria set on idea generation.
I will spare you too much reading.
Disclaimer: Yes, I used Codex to help code and Claude for some verbiage. I am not new to coding but I know there is a huge stigma in the business for this AI coding assistance bits. AI has helped to allow me dump my ideas down to a worker - and not have to take the time to sit and learn to code to get ideas out. Democratized software could have a grand effect.