I can run several agents in parallel.
That does not mean I should let all of them think before I do.
I recognise a pattern in my own work: once agents are researching, implementing, reviewing, and proposing alternatives at the same time, it becomes very easy to spend the day judging answers instead of forming one.
That is efficient. Sometimes it is exactly what I need.
But judgment is partly built in the uncomfortable phase before the options arrive — when I have to decide what matters, which constraint is real, and what I actually believe. Skip that phase often enough and the muscle that produced the opinion quietly becomes the muscle that picks between other people's.
So I increasingly want a few protected surfaces where AI comes second: the first architecture sketch, the first explanation of a difficult trade-off, the blank page before a major decision.
Not everywhere. Most work does not need it, and using agents to argue both sides of a question genuinely improves thinking — as long as I am still the one synthesising.
If AI produces all of the candidate thoughts, choosing among them is not quite the same skill as thinking.