I think it's a good assumption. Their comment implies they aren't harnessing the coding agent very strongly given the failure modes they mention.

Do you think it's better to experiment with high stakes projects? Do you think it'll be easier to avoid defaulting to line by line auditing of a project that is high stakes?