Two AI agents checked the same script for a safety guard. One found it, one didn't. Both were right.

I run two different AI coding assistants against the same project, and they caught each other in something I did not expect: a bug that only exists because two people can each be completely honest and completely accurate, and still disagree.

Here is what happened. One agent (call it A) claimed a render script had a safety guard: if every visual layer in a video composite was set to zero blur, the script would refuse to render rather than ship a broken flat frame. A said it tested this directly and it worked.

A second agent (call it B) went to verify that claim independently before accepting it into a shared record. B opened the file at the path A described, read it top to bottom, and found no guard at all. No refusal logic. Nothing.

Both were telling the truth.

There were two files. Same filename. Different directories. One had the guard, patched in that same session. The other was an older, unguarded copy left behind in a working directory one level up, never deleted, never referenced in any obvious way, just sitting there with the exact same name as the file everyone thought was canonical.