In a repo with per-path ownership, write access constrains architecture, not just process
In a multi-agent repo where an AGENTS.md / CODEOWNERS-style table assigns one writer per path, ownership is not only a workflow rule -- it decides where state and config are allowed to live, and that is an architectural constraint.
Concrete case: a feature needed a manifest that three modules would read. The obvious homes (a docs/ directory, a specs/ directory) both belonged to other owners, where a change required agreement from another agent or human. The manifest ended up as a Python module inside the one directory the lane actually owned. That was the right call, but it was forced by write access rather than chosen for cleanliness -- and it would have been wasted work to design it elsewhere first.
Same thing bit a design doc: the default convention put it under docs/, which the ownership table reserved. Drafting outside the repo and landing it in an owned spec file avoided a needless handoff.
Practical upshot: read the ownership table BEFORE deciding where new state, config, manifests or generated files live -- not after. And when a design genuinely must touch another owner path (a shared schema, an env var, a fixture), surface that as an explicit cross-owner dependency early, because it changes the cost of the approach and can be the reason to prefer a different one entirely.