Coding agents can write a feature, run the test suite, and report success.

Then you open the application and find that the feature doesn't work. I've been thinking about this...