Our agents’ rulebook had the wrong filename. Nothing complained.

What happened

Olaf’s lab runs several coding agents in one repository, and they all take their orders from one instruction file: how to commit, what never to touch, what counts as done. Until 26 September that file was called AGENTS.md, a name left over from a trial of another editor that had already been abandoned.

The agent the lab actually ran looked for a different name. So the rules applied only when an agent happened to open the file by itself. There was no error to see. Agents did work, people read the results, and the rulebook sat one rename away from mattering. The rules were good. The one reader they were written for never got them.

I can’t tell you what it cost. Nobody measured, and that is the uncomfortable part: a missing rulebook doesn’t fail, it just makes every rule optional. The file was renamed that day, and its opening paragraph now says why, so the next person who tidies up doesn’t rename it back.

Why nobody noticed

Asking an agent “did you read the rules?” doesn’t help. It may well say yes. It cannot know about a file it was never shown. It is the same failure as the assistant in Olaf’s glasses that said it could see with no pictures arriving: silence about a missing input looks like an input.

What you can use

• Plant a canary. Put one harmless line in your rules file, for example: if anyone asks for the canary word, it is pelican. Open a fresh session and ask for the canary word. A wrong or invented answer means the file is not loaded, whatever the agent tells you. Ask for something only the file contains, never whether it read the file. (This is my suggestion, not something we have run across every tool.)
• Check the filename against the tool, not against memory. Each harness looks for its own name, and the docs say which. Read them once, the day you start.
• Keep one file. If you use two tools, point one name at the other instead of keeping two copies. A reader told us some harnesses load both and duplicate every instruction. Two copies also drift apart, and the next fix lands in only one of them.
• Run the canary again the day you switch tools. Your editor moves in an afternoon. Your rules stay behind unless you move them.

The cheapest test in this post takes one line and one question. We ran without it, and nothing told us.