Do AI coding agents ever confidently make the wrong assumption about your existing codebase?

Heat trend

Collecting trend data

The percentage is based on available heat signal, not comment count or independent people.

Why it matters

This covers a coding tool or code-capability update — useful for developers assessing workflow changes and reusable value.

For example, assuming an API behaves a certain way, misunderstanding an existing utility/dependency, or getting a business rule wrong.

How do you currently catch these assumptions before the agent makes changes?

I'm specifically interested in the cases where the agent sounds completely confident but is actually wrong.