This analysis is almost bafflingly stupid.

They conflate domain expertise with coding expertise, and then assess that people with domain expertise demonstrate great success at coding tasks, which suggests coding agents are so good at writing code that domain experts can now cut software engineering experts out of the loop entirely. Yet if you look at their classifier, it classifies user expertise almost exclusively according to standards that measure expertise in coding. No wonder it predicts success at coding tasks. This just in: people who know how to develop software are better at developing software. What a fucking joke.

I couldn't quite parse the article headline, perhaps they mean returns like returns on investment?

A more useful headline (that fits in our limit) might have been the takeaway they want you to have:

The more domain expertise brought to Claude sessions, the more end in success