THE SHORT ANSWER
Let the agent make the edits, and make a person responsible for reading them. Nielsen Norman Group's October 2026 study of expert Claude users (Tanner Kohler) found that almost nobody typed into their own context files, that they rarely checked what Claude changed, that none used version control, and that one participant discovered a global file governing what Claude was allowed to do that he had not known existed. I let agents maintain mine, and they are in git, so I counted on October 10, 2026: 11 files, 2,322 lines, and 51 commits since March 30. At least 21 of those commits carry an agent co-author line, and I cannot attribute the other 30 because the trailer was not always added. Twenty-three lines carry my name and a date. Four practices follow: put the library in version control, have the agent stamp any rule a person stated with who and when, read the diff of the global files weekly, and count the files no person has opened. Judging an edit by whether later outputs look fine is the easy check, and a wrong rule passes it for a long time.
The question usually gets asked as agent or human, and the Nielsen Norman Group study settles the typing half: the agent is faster and more consistent, and experts already work that way.
The half it leaves open is review. In that study, edits were judged by later outputs and nobody kept a history. Without a history you can't answer what changed, and without a stamp you can't answer who said so.
The counts from my own repository, including the 30 commits I can't attribute, are in Who Writes Your Agent's Rules? I Counted the Commits. The earlier case of a check that answered the easy question is Every Check Was Green. Four Sources Were Dead.