Review diffs like you'd review a colleague's PR
The habit that matters most when working with an agentic coding tool is the same one that mattered before it existed: read the diff before it ships. An agent can write plausible-looking code that's subtly wrong — an off-by-one in a loop it added, a fix that handles the reported case but not the adjacent one, an edit that satisfies the letter of your instruction while missing its intent. None of that is a sign the tool failed; it's a sign that generated code, like any code, needs a human check before it's trusted, exactly the same standard you'd hold a colleague's pull request to.
This is also why small, reviewable diffs beat large sprawling ones. A change touching three files with a clear purpose is something you can actually verify carefully in a minute or two. A change touching forty files because a task's scope grew unchecked invites exactly the skimming that lets real mistakes slip through — which loops back to why tight task scoping and plan review from earlier lessons matter in the first place.