I read “in their own words” as being equivalent to know what they are doing. It’s way to easy to ask an agent “Hey, fix this issue for me” without understanding what happens. And that’s something we don’t want. If you are able to explain what the agent did and why (without asking it), then you are likely able to write the same code.
My understanding is that we don’t want contributors who simply, without any effort, create PRs that they think is their work. For me it’d be inacceptable because in the end, they are not the contributor, it was the agent who contributed. If they know what the agent is doing, if they guide it properly and if it can be simply considered a way to generate code that we would anyway generate on our own, then it’s ok. If it’s just making the agent work instead of you and trust it, then it’s not ok. The important part of what Serhiy said is (for me):
Not everyone is an English native speaker and I’d rather have someone using an LLM for translation (which, btw, is accepted) than a message I can’t read because it’s written in the author’s native language (which may be different as well and not understood by anyone but the author!). My PR commit messages are very poor in general if I’m writing an incremental PR (it’s usually “first commit message is good” and then it’s “fixup”, “??”, etc). The message I use for the commit, though, is either the first commit message that was already well-written (usually it’s similar to the NEWS entry), or something I spend more than 10 minutes on it just to find the right formulation without exposing too much details when they are not needed (while keeping enough detailed information about the implementation when it matters).
If I can automate that process and simply edit the agent’s suggestion, I’m also fine.
It doesn’t work for every topic. New contributors can easily tackle simple issues but even amongst core devs, there are hardly any experts in turtle, IDLE and imaplib. Contributing to imaplib requires to read and understand RFCs as well as knowing the existing infrastructure, and asking new contributors to learn our process by contributing to such components is honestly not the best approach IMO.
I wished we had “Generated by” instead. Because, TBH, making the agent the co-author when 90% has been generated by it is a misattribution. I am the co-author, and the agent is the main author in this case. Having “Reviewed by […]” as well would be great (I think the Linux kernel also has different attributions depending on who does what).