A coding agent that reports "done" is making a claim, not handing over proof. The tests it ran might not cover the change. The file it says it edited might be a different copy. The command it says passed might have been piped through something that swallowed the exit code.
Most people who work with these tools end up with some habit for this. Some read the whole diff. Some re-run the tests themselves. Some only trust a result that a second, separate check produced. Others accept "done" for small changes and only dig in when something touches money, data, or anything that is hard to undo.
We are curious where you draw the line:
- What is the one thing you always check before you accept an agent's "done"?
- Is there a kind of change where you have stopped checking, because it was never wrong?
- Have you been burned by a "done" that was not done? What gave it away?
Short answers are welcome. We are collecting the habits that actually catch problems, not the ideal process, so please share yours in the comments below.
