This was a hackathon project by me and my friend Vish Rajiv. It won us the Angellist Agents hackathon. The goal was simply to find an interesting form of an LLM agent--something that uses an LLM's output to determine its next input, and is thus able to run indefinitely.
Whybot builds a recursive knowledge graph around a seed question. You seed it with a question, it answers it, then asks follow up questions, answers those, and on. It's good for getting a really deep dive into a topic.
Making a parallel, streaming breadth-first-search that you can pause and resume was more tricky than I anticipated.