This week brought big changes to AI coding tools, with companies racing to build smarter agents that can write and review code. Anthropic released Cowork, a Claude agent that works directly in your files without needing any coding knowledge. Meanwhile, xAI is rebuilding its coding assistant after falling behind competitors like Claude Code and OpenAI's tools.

However, the rush to use AI for coding is causing problems. A new study found that when developers use AI coding tools very often, their companies experience more bugs and crashes. About 96% of frequent AI users work nights and weekends because problems keep popping up after the AI writes code. The issue is that while AI can write code super fast, the rest of the system cannot keep up with testing and checking it for safety.

The good news is that AI is getting smarter. Researchers at USC found that GPT-5 improved from 39% to 96% success when given feedback about mistakes, even when learning a language it barely saw during training. This shows AI can learn and get better, not just use what it already knows.

The challenge now is balancing speed with safety. Companies want AI to write code quickly, but they also need to make sure the code works properly and does not break things. The solution is pairing AI tools with strong human judgment and careful checking, not just letting AI do everything.