Picture a website that wants AI agents to find its knowledge files (descriptions, data exports, instructions for agents) without crawling around. A first-revision IETF draft, dated 30 September 2026 and written by the independent author A. N. Besleaga, proposes one fixed place for that list, and a fingerprint on each file so an agent can tell whether it got the right bytes. It is a first draft with one author and, as far as I could see, one implementation. Watch. If you publish files for agents to fetch, copy the fingerprint idea; do not adopt this format yet.
What it is
A site serves a single JSON document at /.well-known/knowledge-linkset, over HTTPS, one per site. It is a "link set", an existing web format for lists of typed links, so the draft invents no new media type. Each entry points at a file (a graph export, a context file, a text file for agents, a change ledger) and may carry a fingerprint: a SHA-256 hash (a short code that changes if even one byte changes) written in the standard HTTP digest syntax. A link without a fingerprint counts as unverified, not invalid.
What the fingerprint proves, and what it does not
The draft is careful here. A fingerprint catches truncation, stale caches, mirrors that alter bytes and plain corruption, which matters once agents fetch instructions and data automatically. But in its own words a digest "proves only that an artefact is what was published". It does not say who published it: anyone who controls the site can make the list and the files agree. No fingerprint makes content safe to follow.
What it leaves out
The draft defines no agent identity, no way to authenticate whoever fetches the files, no transport between agents, and no format for the files themselves. It also says publishing the list "grants no rights and no permissions of any kind". The files and their meaning come from a separate specification and engine published by the same author. The only implementations I found are that author's own: a reference site, a validator and test vectors, tied to one release of that engine. I did not search exhaustively for others. The draft expires on 3 April 2027.
What to do
Watch. The useful part is small and easy to borrow: if agents download your files, publish a hash next to each one so a damaged or stale copy is noticed. Do not wait for, or depend on, this particular format.