The agents did all the work but _somebody_ has to test it for real on their own data to find the edge cases overlooked by AI. That's what users are for nowadays.