Hi there,

Most of what I learned about reviewing code over the last ten years was learning to read the writer, not the diff. I watched for where they hesitated, where they left a comment that didn't explain the code but explained the decision, the TODO they hadn't cleaned up, the tempFix variable from two months ago that nobody had touched. That stuff told me where to look.

Agent PRs don't have any of it.

Every variable has a real name, every function looks intentional, every abstraction has a plausible reason to exist. Nothing is marked tentative because the agent isn't tentative, it chose, it moved on. And the line where it picked something nobody on the team would have picked reads no different from the 50 lines of perfectly fine code around it.

So my old review instincts started missing things. Not on the obvious cases. The agent doesn't add a new database write in a place I'd notice. It introduces a plausible abstraction in a file I wasn't planning to read carefully. The abstraction quietly changes what a downstream function returns, but only on one branch I hadn't thought to check.

I had to stop reviewing in the order I'd always reviewed in. I had to stop using "this section looks confident" as a signal that meant "this section is fine". That signal was always doing more work than I'd noticed. It told me where the writer was paying attention, so I could pay attention there too. With an agent, confidence is the default state of the entire file.

That's what the workshop on June 24 is for. Ninety minutes on how to triage a diff when the legibility cues you've been reading for a decade are gone. I'll walk through the failure modes I keep catching, the order I run a review in before I look at any code, and how to bake the whole thing into a workflow that doesn't slide back into vibes the third time you're reviewing agent output that day.