Why does a fixed agent bug come back weeks later?

A fixed bug comes back when nothing keeps checking for it after the fix ships. It’s easy to verify a fix against a hand-written test that describes the bug in general terms, merge, and move on, while the actual failing trace, the one with the specific message history and tool outputs that triggered it, never becomes a permanent check. Weeks later, a different change reintroduces the same condition, and nothing catches it, because no test was built from the original case. The fix that holds is the one where the replayed failure becomes a regression case: the exact trace, kept, and run against every future change. Keeping that case as a standing check is what stops the bug from coming back.