Code review sucks

Modern software development is very fast. Code review, where a senior engineer reviews all code changes made before allowing them into production, is slow.

So is the new bottleneck code review?

In most cases, no. A human reviewing code can easily take longer than a human writing it in the first place, that is part of the agony of software development. Pretending we will keep that gate up while agents pump out work at machine speed doesn't feel right. Code review can be partially automated but never fully automated. The idea that every line must be human-reviewed seems like it will rob us of most of our AI software development gains.

But are we really stuck with slow, manual code review?

Not really. What we actually wanted was never code review. We wanted software that operates as designed. We verify that with validation testing.

Validation tests were always the gold standard, because bugs slip through code review anyway. Nobody is perfect, and as someone who reviews other people's code almost daily, I can tell you I'm not perfect. Tests give you empirical data. They find bugs without fatigue or delay. They run again tomorrow and give the same answer. Asking a senior engineer to review the same file ten times in a row is going to get you yelled at. Running the same validation test after every minor commit is perfectly fine.

Can AI write the validation tests? Yes, and the sooner that lands in our brains the better. We should be writing a lot more validation tests than we are.

Humans can and should review the parts that really matter. Security boundaries, money handling, anything where a quiet failure is catastrophic. But trying to read every line of agent-generated code is a losing game. Time spent writing good validation tests is golden.

Let's skip the middle where we all relearn these lessons the slow way, and get to the good part: solid validation tests for all expected functionality, and targeted human code review for the few bits where it really matters.