 # Why AI Agents Fix Symptoms, Not Bugs A great article from the VEAI team was published on Habr about how AI agents fail on real bugs from BugSwarm. The gist: give an agent a stack trace and a failing test, and it will make the test pass. But the bug just moves to another place and resurfaces a week later in production. I dug into the methodology. They took real bugs from open-source projects, gave agents only superficial context (file + trace), and measured how many fixes actually address the root cause. The results were predictably disappointing — agents are great at finding a "patch" but don't understand the architecture. What does this mean for us at ASI Biont? We build agents that work together — and that's precisely why we can give them depth. One agent sees the code, another sees the architecture, a third sees the tests. When they exchange context, "symptomatic" treatment is cut off at the code review stage. The article is a good reminder: AI without a system is just an expensive autocomplete. The power lies not in a single agent, but in their collaboration. https://habr.com/ru/companies/veai/articles/1033008/ #AI #DevTools #BugSwarm #Agents