AI-generated code looks finished until you have to change it
I use AI agents to build real products. They can produce code, tests, documentation, migration plans, and a convincing completion message in one run. The speed is real. The new failure mode appears