Most PMs are still prompting agents. The job is shifting to designing the loops that do it for you.
Loop engineering means you design a system that finds the work, hands it to an agent, checks the output, and decides what's next. You define the goal and the stopping condition. Your job shifts from operator to architect.
You stop crafting prompts for every task and start designing the environment the agent operates inside: the context it starts with, the verification criteria it runs against, the memory it carries between sessions, and the evals that tell you whether the output actually met the bar you set.
The catch: a loop running unattended is also a loop making mistakes unattended. The PM who uses this to move faster on work they understand deeply gets compounding returns. The one who uses it to avoid understanding the work gets compounding debt. The loop doesn't know the difference, you do