In this paper, we develop a reinforcement learning agent that frequently finds the minimal sequence of unwinding crossovers for knot pictures with up to 200 intersections. This provides an upper bound on the number of unwindings for 57,000 knots. The pictures of the connected-sum knots are taken as having opposite signatures, and the sums are superimposed. The agent finds examples of multiple unwinding crossovers in the set of unwinding crossovers that produce superparabolic knots. Based on this, we show that given knots K and K' satisfying some weak assumptions, there exist their connected-sum pictures and u(K) + u(K') unwinding crossovers, such that changing one of them produces a prime knot. As a byproduct, we obtain a dataset of 2.6 million distinct difficult-to-unwind knot pictures, most of which have less than 35 intersections. Assuming countability of unwindings, we determine the number of unwindings for 43 knots with up to 12 intersections, where the number of unwindings is unknown.