Blog
16 June 2026Sahil SapraFaster Than We Can Test

Convincing is what slips through review

AI-written code can pass PR review and still take down production. At review, 94% rated it higher quality than human-written code. In production, 82% hit a production failure within six months. Source: New Relic, 2026 State of AI Coding.

The best code shipped all quarter took down the checkout page at 2am. It had passed pull request review in twenty minutes, with no comments on it.

This is the reality for many engineering teams today. The code looks right and reads right. There is no typo or missing import to catch. The logic is quietly wrong, in a way that only shows under real load, on a path nobody thought to test.

That is the point. AI wrote convincing code, and convincing code is exactly what gets through a review.

A New Relic report published on 10 June put numbers on it. 94% of engineering leaders said AI-generated code was higher quality than what their own developers write, judged at the moment of review. Then it shipped. 82% had a production failure tied to AI code in the previous six months, and at least a quarter of the code needed serious rework. The gap has started to be called "agent debt."

Here is the trap. A pull request review measures how code reads. Production measures whether it is true. AI is very good at the first and indifferent to the second, and the space between them is where the 2am incidents happen. We treated a green checkmark as proof when it was an impression.

So when AI hands you a pull request that looks perfect, what stands between "looks right" and "is right" on your team today?

The data behind this post

  • New Relic's 2026 State of AI Coding report, published 10 June: the press release: newrelic.com
  • New Relic's 2026 State of AI Coding report: the full report: newrelic.com

Build today. Know it's safe by morning.

7:30 am. Ready to ship.