it's been long demonstrated across every domain that fast feedback loops are the #1 thing to optimize for i think the big llm labs are getting too obsessed with long running async agents sonnet making all these intermediary markdown files has made it unusable