AI can now diagnose why software builds fail intermittently — using just a dozen examples

Researchers built a system that reads software build logs and automatically identifies why jobs failed intermittently (flaky tests, network outages, resource problems) with 84% accuracy, needing only 12 labeled examples per failure type. This means developers spend less time manually diagnosing broken builds and can redirect that labor to actual coding.

Software developers waste enormous time chasing build failures that aren't their fault — network hiccups, infrastructure glitches, resource exhaustion. The system addresses the second half of the problem: not just detecting that a failure happened, but diagnosing its cause fast enough that a human doesn't have to. This matters because diagnosis time compounds — it interrupts the developer, often requires calling in a specialized ops team, and delays the whole pipeline. The technique also works with minimal training data (12 examples per category), which means companies can deploy it immediately on their own logs without massive annotation effort.