A shared filename is not a shared contract

Near the top of the radar’s AI feed this week: a feature request on the Claude Code repository asking it to support AGENTS.md. Three hundred sixty one points, two hundred eighteen comments. That is a lot of energy for a request that one harness read another harness’s instruction file.

Scroll the same day and you can see why people care. A launch for an open source sandboxed agent harness for teams. A tiny open source coding agent from Vercel. One vendor’s agents moving into other people’s editors. Nobody reading that list is going to end up with only one harness, and most of them arrive wanting their own file at the root of your repo.

The easy read is that this is a standards squabble and somebody should just pick a name. I do not think the filename is the hard part.

The vault’s cross-tool notes describe the real failure: without a shared work surface, the person becomes the hallway between systems, carrying context by hand from one private chat to the next. Its handoff rules are specific. Put the outcome, source files, constraints, and definition of done in the shared task. Make the receiving agent’s context self-contained. Require receipts. Keep merge authority explicit, because a handoff is not an approval.

Now hold that list next to a markdown file sitting in a repo root. The file can carry the repo shape, the commands, the constraints, the house style. It cannot carry what the last agent decided, what it already tried, what it broke on the way, or whether a human ever signed off.

Agreeing on one filename would be a real improvement and I hope it happens. It would also fix the smallest part of this. The instructions were never the thing that got lost in the handoff.