AI Coding Agents Broke My Code Review Habit, and I'm Not Sure That's Bad
I used to read every diff like it owed me money. Now an agent pre-reviews and I read the interesting 10%. The skill didn't disappear, it moved.
For years my rule was simple: no code ships until I've read every line of the diff. It was slow, it was a little self-important, and it caught real bugs. Then I hired an agent to pre-review, and the habit quietly fell apart.
What changed wasn't trust, it was triage. The agent now reads the boring 90%, the renamed variables, the moved functions, the import cleanup, and flags the three lines where something actually has consequences. I read those. My review time dropped from an hour to ten minutes and I'm fairly sure I catch more of the things that matter.
The fear everyone voices is that I'll stop understanding the codebase. I think that's backwards. Understanding was never in the reading of boilerplate. It was in the argument about why a change is shaped the way it is, and I still have that argument, just about fewer lines.
There's a cost I won't romanticize. When the agent is wrong about what's boring, a real change hides in the 90% and I miss it. So I keep one habit: every Friday I read one full diff end to end, a random one, to remind myself what the boring parts look like. It's a calibration, not a ritual.
The skill didn't vanish, it relocated. I'm no longer a human linter. I'm the person who decides which 10% is worth a human, which is a different and probably better job than the one I was doing.