I've read plenty of anecdotal evidence of people finding their diligence declining over extended periods of agentic coding.
Then from the other side, just because an agentic system can be used to create a functioning system, I'm not convinced they can own it for any length of time. The problem as I see it is thinking of the LLMs interactions with a codebase in information theoretical terms; each interaction an LLM has with a codebase will introduce noise to the signal.