The SDLC isn't dead but the review queue might be
The human stays on intent and everything else goes into automated gates, but a regulated shop still needs something for the auditor in place of the approval click.
You want to get attention, you don’t say things are changing. You just say this is dead.
“Dead” refers to Boris Tane’s post The Software Development Lifecycle Is Dead, which argues that coding agents collapse the classic stages of the lifecycle into a single tight loop. There’s a good discussion of the post from both sides on the Changelog & Friends episode. I think the title is overstated, but the underlying argument is sound.
You’re assuming each of these is a massive undertaking. I always felt like it should be smaller, tighter, and loopier.
I always thought so.. the hand-offs between stages were the real problem.
We’ve ritualized and held sacred these steps along the way because we’ve had to and that’s been our comfortability. So as humans, we have this comfortability gate.
A lot of code review is precisely such a gate. It’s a way for the reviewer to gain confidence that someone looked at the change. That’s not the same as proving the code does what it’s supposed to do.
If you have an agent that files a few hundred PRs per day, and your team can review ten, your reviewers don’t suddenly become more careful. Instead, you get a backlog, and approvals degrade into clicks to clear the pile.
You’ll always care about some version of code quality, but the purpose of caring about code quality is, is the intent met? Does it pass tests? Does it actually work in development environments? Does it pass the smoke test where we spin it up 50 times and battle test it?
The human stays on question one, is the intent satisfied, because we need human judgment about what was really being asked for. Everything else gets automated into gates that every change has to pass through: run the test suite, spin it up repeatedly and do smoke runs.
I’m not sure exactly how, but I’m thinking code quality over code review. That’s where I’m leaning.
I lean the same direction, but there’s one prerequisite hurdle. In regulated organizations, the second-person approval doubles as audit evidence that a non-author verified the change before release. So the open questions are: what would an auditor accept in lieu of the approval click, and can a quality gate generate it?