AI coding agents are getting better at writing code.
But I think we are approaching a more difficult question:
How do we know that an AI agent made the right engineering decision for the current state of a software system?
Passing tests is important.
But passing tests alone does not necessarily tell us whether an agent understood:






