The project is turning out to be way harder than I thought (which is a good thing, means there’s something interesting in there). Some challenges so far

- writing is very personal and subjective; LLMs are not very good at it. need lots of guardrails and guardrails to emerge on the fly - supporting fluid transitions between agent taking the lead and human taking the lead is not easy - nontrivial concurrency issues - keeping the user in flow with a bunch of agent activity happening - parallelizing work such that human think time and agent think time happen at the same time, not sequentially (ie in chat based products)