Researchers Find a Simple Fix for Broken Code — and It Isn't AI

Software is built from reusable pieces of code called "dependencies" (ready-made building blocks other programmers wrote). Each block comes in many versions, and they don't all get along. Ask for version 2 of one block and version 3 of another, and the whole program can simply refuse to run. Programmers call this "dependency hell," and it eats hours of their week — time that could go into fixing bugs or building features you actually use.

A team of researchers built a tool called PLLM+ to attack the problem. Their approach is deliberately unglamorous. First, it tries cheap, predictable steps: figure out which version of Python the code expects, look up whether this exact combination has worked before, and double-check that the packages exist online. Only if all that fails does it call in AI (specifically, two AI "agents" — programs that take actions on their own — one proposing fixes and one criticizing them).

The results were striking. Tested on 2,891 code samples that were known to be broken, the tool fixed 1,500 of them — up from 1,169 for the earlier version. And 1,495 of those 1,500 fixes came from simply replaying previously working combinations. The AI contributed just 5. Average repair time dropped from about 369 seconds to 72 seconds, a roughly five-fold speedup.

The lesson reaches beyond code. When a boring, checkable method works, it beats a clever one that guesses. The researchers themselves frame AI as a backup, not the hero. For you, the practical upside is subtler but real: fewer hours burned on setup means software teams ship features, security patches, and bug fixes sooner — and waste less of the money paying for all that lost time.