My main takeaway from that post was that, apart from "don't use assembly", none of the experiments showed strong evidence either way, and the things that took time were when the agent got stuck doing silly things like invoking cargo with the wrong arguments. Ie. there's a lot of measurement noise, and not any strong signal, and "LLMs are good at generating $PL" isn't well-founded as an argument for picking $PL. So we're back to the old arguments for picking a language, like "I find it readable/enjoyable"