Code Review Used to Work. Then the Agents Arrived. | by Jaspreet Singh | Jun, 2026 | MediumSitemapOpen in appSign up<br>Sign in

Medium Logo

Get app<br>Write

Search

Sign up<br>Sign in

Press enter or click to view image in full size

Code Review Used to Work. Then the Agents Arrived.

Jaspreet Singh

8 min read·<br>Just now

Listen

Share

After 13 years shipping mobile apps to tens of millions of users, I watched good engineers get buried under 45 PRs every two weeks. Code review didn’t break it just stopped scaling. So I fixed that.

Every company wants 3x, 5x, 10x productivity. AI coding tools promised exactly that and they delivered. Output went up. Feature velocity went up. The number of PRs ready for review every sprint went up.<br>Nobody planned for what that does to the engineer on the other end of the queue.<br>AI writes code at machine speed. Humans review it at human speed. That gap is the problem nobody is talking about loudly enough.<br>I’ve spent 13 years in mobile engineering, and I’ve never seen engineers more burned out by code review than right now. Not because they’re less capable. Because the queue that used to have 15 PRs before a release now has 45 and the expectation is still that every one gets a thorough review before it ships. So engineers push through. PR after PR, afternoon after afternoon, trying to hold an ever-growing codebase in their head while the backlog keeps climbing behind them.<br>And bugs slip through. Not because anyone was careless because the sixth PR of the afternoon, a 400-line diff with a stale useEffect ref dependency, looks fine when you're tired and there are still twelve more behind it.<br>I’ve seen what that does to engineers. The guilt. The “I should have caught that.” The slow, quiet demotivation of doing everything right and still watching things leak into production.<br>The output problem is solved. The review problem is not. That’s what I built Preflight to fix.

Press enter or click to view image in full size

The Acceleration Nobody Warned You About<br>Here’s what actually happened when AI coding tools went mainstream.<br>Code output went up. Way up. Pull requests multiplied. The diffs got longer. The context got harder to hold. And human reviewers whose reading speed hasn’t changed since we started staring at screens quietly became the bottleneck nobody wanted to talk about.<br>A 2026 Faros AI study of 22,000 developers found that as teams adopted AI tools, the per-developer defect rate jumped from 9% to 54%. Code churn went up 861%. Review time went up 441%. And PRs merged with zero review rose 31% not because teams stopped caring, but because reviewers simply could not keep pace with the volume.<br>Nobody decided to stop reviewing. The code just arrived faster than review could absorb it.<br>I watched this happen in real time. Not on charts in my own PRs, my own queue, my own increasingly inadequate afternoon review sessions. I was the mobile lead. Every diff that shipped wrong was on me. And the tools I had linters, static analyzers, CI checks were built for a world where the bug lived in a type mismatch or a missing null check.<br>The new bugs were different. Logical. Contextual. The kind a second human would catch, if that human wasn’t already reviewing seven other things.<br>So I built a second human.

Introducing Preflight<br>Preflight is an AI code reviewer that runs on every PR before a human lays eyes on it.<br>Not another linter. Not a static analyzer. A reviewer one that reads the diff the way a senior engineer would, asks the same questions a senior engineer would ask, and posts its findings directly on your PR.<br>One command:<br>git diff main...HEAD | python3 preflight.py -That’s it. Preflight reads the diff, loads the right rule set for your stack (iOS, Android, TypeScript, Python, Go), runs two independent model passes, verifies every finding against the actual code, and posts a PR comment with only the things worth saying.<br>No noise. No false positives that cry wolf. No findings that cite lines that don’t exist.

Press enter or click to view image in full size

The Architecture That Makes It Different<br>Most AI code review tools are wrappers. They pipe your diff into a model, get a response, and display it. Simple. Fast. And deeply unreliable — because language models hallucinate, and a finding that cites a line that isn’t there is worse than no finding at all.<br>Preflight is built around a different idea: chain-of-custody verification .<br>Here’s how it works:<br>1. Stack detection. Preflight reads the diff’s file extensions to automatically load the right rule set. A Swift file gets iOS rules. A .kt file gets Android rules. You don't configure anything.<br>2. Two independent passes. The model reviews the diff twice, independently. This isn’t redundancy, it’s recall. The two passes surface different things. The merge step at the end is what eliminates duplicates.<br>3. The verifier loop. Every finding must cite a specific rule ID and quote verbatim evidence from the actual diff. If a finding can’t point to...