Why Humans Can't Keep Up With AI-Written Code
summary
Agents produce 2,450% more lines of code but releases rise only ~30%, and five failure modes matter: humans agree with wrong AI 4 out of 5 times, agents fail on very large codebases (accuracy collapses past ~50k LOC), errors compound over long task chains, LLM judges flip answers 7–34% of the time, and exploit windows are near zero. Sonar’s approach uses an intent+guide rails+verification+evidence formula and products like SonarVortex and SonarQube to cut token costs 36% and commit issues 92%.
AI-written summary · about
chapters (12)
- 0:00Intro
- 0:26The ball bearing that changed car factories
- 2:23Software already borrowed these ideas
- 2:55Agents break the factory
- 5:19Problem 1: Human in the loop doesn't scale
- 6:45Problem 2: Agents struggle with big codebases
- 8:04Problem 3: Errors compound
- 9:06Problem 4: LLM judges are unreliable
- 10:03Problem 5: Time to exploit is near zero
- 10:25The code confidence formula
- 12:36Results from customers
- 15:49Code is abundant, confidence is scarce
