AI Can Help Fix Software Bugs — But the Cheap Version Works Best

Every night, software companies run automatic checks — called tests — to make sure their products still work. When one fails, a human engineer has to dig through piles of confusing logs (computer records) to figure out what went wrong. It's slow, tedious, and happens constantly. A Swedish networking company called Westermo teamed up with researchers to see whether AI could do the first pass of that detective work.

They built two versions of the same helper. One was a single AI that reads text and points to a likely cause. The other was a team of AI 'agents' — programs that can take steps on their own, like looking things up and checking each other's work. Both got access to the same test data and logs. Six engineers then reviewed the AI's reports from two real failures and rated them on accuracy, reasoning, clarity, usefulness and trust. The team also ran the systems 120 times to measure speed, cost and consistency.

The result was surprisingly boring, in a good way. Neither version was consistently better in the eyes of the engineers. But the simple single AI produced its reports faster and for less money — which matters, because every extra AI step burns time and computing costs. For routine bug-hunting, the researchers concluded the plain version is the practical choice for now. The fancier agent setups may still pay off, but only on much harder problems.

What does this mean for you? Software bugs and outages cost companies billions and delay the apps and services you rely on. If AI can do the boring first pass of diagnosis, engineers fix things faster and spend their time on harder work. And if the cheaper tool works just as well, companies don't have to pass big AI bills onto customers. One caveat: this was a single company, two failure cases and six reviewers — a small experiment, not a final verdict.