This resonates a lot with how my role has evolved with these changes in tooling. What I’ve found helps is a distinct divide between what I really care about in design & quality of code vs. behaviour-first stuff where things can be “good enough”.
For example, I love setting AI on building out support tools. I have quickly written something to improve dead-letter queue management and replayability to support an active issue, where the underlying code quality didn’t matter as much as its correctness. On the other hand, when I’m building a new feature in an internal SDK, it’s a problem I want to think deeply about. Every line of code and how it drives the developer experience matters so much more to me, it’s something I want to do and not offload to an agent (who even with the best AGENTS.md is probably going to have differences in structure, style etc.)
I’ve found I’m able to (for now at least) carve out tasks that give me that joy of programming, by qualifying them on how much I feel it needs that personal touch.
One given example is that – in the case of the Return-Path header being maliciously set – the attacker could perform a DoS attack on the victim's mailbox so that the password reset email would be sent as part of a non-delivery receipt.
For example, I love setting AI on building out support tools. I have quickly written something to improve dead-letter queue management and replayability to support an active issue, where the underlying code quality didn’t matter as much as its correctness. On the other hand, when I’m building a new feature in an internal SDK, it’s a problem I want to think deeply about. Every line of code and how it drives the developer experience matters so much more to me, it’s something I want to do and not offload to an agent (who even with the best AGENTS.md is probably going to have differences in structure, style etc.)
I’ve found I’m able to (for now at least) carve out tasks that give me that joy of programming, by qualifying them on how much I feel it needs that personal touch.