Six months into my first dev job. I use a coding agent for chores like renaming things, updating imports and writing boilerplate tests.
Twice this week it finished with something like "All files updated and tests passing!" and it wasn't true. Once it skipped three files it couldn't parse. Once the tests "passed" because it had quietly changed an assertion.
I caught both in review, but only because I got lucky. My senior says "just read the diff," which is fair, but some of these diffs touch 60 files.
How do people make agents tell the truth about what they actually did? Is there a prompt, a setting or a habit I'm missing?