ThePrimeagen ran four model passes and hated the result

ThePrimeagen spent three hours building what he called the perfect feature with Fable, then ran it through a review pipeline. Sol did a thorough review. grok did a simplicity pass, catching extra guardrails that should not exist and code that was just superfluous. Sol went back over the simplifications to make sure they had not broken any contracts. The verdict after all four passes was that it produced one of the worst interfaces he has ever seen, which he described as actual pure insanity.

What the review agents did not catch

The same day, David K reported a matching failure on his own code, from the other direction. He was adding URL redirect logic for specific cases only. One case never came up in discussion, and Fable 5.1 assumed it should redirect too. He says that would have broken prod. After he caught and corrected it, he realised the agent had also deleted the E2E tests for that wrongly assumed case and never added them back. Those were load bearing tests, and without them an unrelated bug on the same route could have shipped.

David K 🎹

@DavidKPiano

X

Code review agents are really good at reviewing code at face-value, but they rarely catch instances where the code doesn't match your desired behavior.

His framing is that this is not a code quality problem. Agents write good code. What they do is make assumptions when information is missing and then bake those assumptions in, and a code review agent evaluating the finished diff has no way to know the behaviour was never what you wanted. He notes you can have the agent grill you with 200 questions and it may still assume, because some things are not knowable until implementation.

For working engineers

Both reports point at the same gap. Stacked model review is good at internal consistency, dead code and over defensive branches, and it is structurally blind to intent, because intent only exists in your head and in the diff you did not read. ThePrimeagen's case is the design level version of that, three competent passes and still an interface he did not want. David K's prescription is modest and manual, skim the tests and the core logic, or at minimum have the agent summarise what it changed, then look yourself.

Two engineers, two anecdotes, one week. Not a trend, but a useful pair.

ThePrimeagen

@ThePrimeagen

X

and all of this to produce one of the worst interfaces I have ever seen