LLMs are now good enough that human source code review has become the bottleneck.
In his latest video, Mo Bitar talks about his experience letting Opus 5.5 build a complicated feature: real-time multi-user support for his app. A few months ago, this would not have been possible with LLMs. Now the feature worked flawlessly after a few hours without his intervention. For a classical engineering team, that would have taken weeks, maybe months to ship.
https://www.youtube.com/embed/2zLuYU_Ub_0
LLMs are adversarial towards their own conclusions
I see the same with Opus 5.5. It doesn't only write good feature code. More importantly, it is really, really good at finding edge cases and everything that can go wrong, because by now LLMs have been trained to be adversarial towards their own conclusions in order to reduce the mistakes they are making (and will keep making). And because of that, they spend much more time analyzing what could go wrong than I would.
Human review makes the code less robust
I work on projects with and without human reviewers, and I add similarly complex features to both. Without a reviewer, I let the LLM build defensive code in depth, and it detects and prevents more edge cases. With a reviewer, I often find myself holding back from letting the LLM add more defensive code for unlikely edge cases, because that code is often larger than the happy path, and every line of it adds cognitive load for the human reviewer.
Human code reviews in the age of agentic development are putting the brakes on development speed. And because the number of lines a human can review is limited, they also make the code less robust than it could be.
Review the result, not the code
While human source code reviews have become obsolete, it is still important to review whether the implemented solution is fit for the task it aims to solve. The time we would have spent reading lines of code, we can now spend trying out the new feature. And we can use human creativity to find ways to improve it: especially in frontend projects, I still find that the look and feel often leaves much to be desired.
Mo makes the same point: a CEO doesn't review the work their employees are doing. They measure the result against the bottom line.
Let the LLM write the code. Let humans judge the result.