💭 It might help to get a very solid understanding of ways and situations in which an llm will respond unlike a human. For long term coherence, I think these issues compound on each other. For coding for example, llm tends towards more complicated answers, adds more than subtracts, and doesn't give up, which scales into increasingly horrifically broken environments. What are the deviations from human behavior for general agency? Maybe allowing the llm to give up is a key step. Maybe it needs some kind of objective tree, but I've avoided that because objectives are usually inherently temporary and temporary fact tracking is another pain point.