Boris Cherny, who runs Claude Code at Anthropic, put it plainly: he does not prompt Claude anymore, he has loops running that prompt Claude and figure out what to do. his job is to write loops.
a loop is a recursive goal. you define a purpose, the agent iterates against it, and the loop keeps running until a real stopping condition is met. the agent forgets everything between runs. the loop does not. that single fact is the entire architecture.