Devin probably uses some Reward Model & Best-of-N to search for good action, which might help explain why it is so slow