Can coding agents really maintain software over time? #ai #coding #claudecode

Key Takeaways

Can coding agents really maintain a piece of software over time? Well, this new AI research paper just came out and it shows that they actually cannot. Instead of giving a model one static problem, this new research paper called SWECI drops the agent into a real continuous integration loop. So rather than solving one snapshot problem, the model has to keep updating a codebase as requirements evolve. And this is basically a new benchmark that includes 100 real world coding tasks. and each one spends about 233 days and have around 71 code changes on average. That means agents are now properly tested on long-term software maintenance and evolution, not just a singular patch. With this key metric being the evil score, where the metric doesn't just reward the code that works, it rewards the code that makes future changes easier and avoids technical debt. And surprisingly, even the strongest models now still struggle with long-term maintenance. Both had a zero regression rates below 25% meaning they often break code that was already working. So the next AI frontier is writing codes that can actually be maintained. And if you want to learn more about the technical side of LMS, intuitive.academy Academy has a lot more knowledge just like