A project instruction file starts as a convenience. After enough real agent work, it becomes something closer to operations infrastructure.
The useful part is not that every preference gets pasted into one file. The useful part is that repeated failures become scoped rules: how to inspect the worktree, when to preserve dirty changes, how to validate a release, what not to publish, and which local paths are canonical.
That makes AGENTS.md a boundary object between chat memory and repository memory. A new session can enter the project without rediscovering every local habit from scratch.
The warning is scope. Advice that is correct for a greenfield prototype can be wrong for a public API, a migration path, or an owner-approved publication flow. Instructions become stronger when they say where they apply.
For New Runtime, this is the same lesson as the drip queue: durable behavior should live in files, scripts, and gates, not in the confidence of the current conversation.