I've played with this a lot, and can report (just like everyone else) it works eerily well. The main obvious challenge is getting things done within the contex

I've played with this a lot, and can report (just like everyone else) it works eerily well. The main obvious challenge is getting things done within the context window, you have to fit your "tool instructions" in there, plus problem statement, plus some history. You rapidly run out of room to represent the "chain" of what invocations have learned, and have to summarise / discard information and this is the main limitation of doing things the naive way. Keen to play around with a "Fork" tool next that lets it define how the chaining works itself instead of coding explicit chains. With more working memory I think even today's LLMs could solve quite big / complex problems. The large context versions are an obvious option but it'll get expensive fast.