The Agent Citation Stack

The questionHow should AI agents cite sources so machines can verify them? Citing isn’t just links. It’s stable URLs, machine-readable twins, and a path from public references to private detail.

Footnotes exist for a reason.

They let a reader separate “this is the author’s voice” from “this is where the claim came from.”

Agents need the same separation, but in a form machines can use.

The mistake most teams make is thinking citation is just attaching a link to an answer. That’s the shallowest layer.

The real citation stack has a few layers:

  1. A stable canonical URL: something a human can click and understand.
  2. A machine-readable twin: something a system can parse without scraping.
  3. Explicit time: `as_of` and ranges, so the citation has temporal meaning.
  4. Addressable examples: claim IDs or other stable handles, so evidence isn’t just “somewhere on the page.”
  5. A private resolution path: if you need deeper detail, you should be able to retrieve the exact object behind the public reference through a paid interface.

That last layer is the one most people ignore.

Without it, a public citation is a dead end. It’s a marketing page.

With it, the public layer becomes a doorway: you can cite publicly, then resolve privately when you need to act.

That is the boundary Hanging Context and Synorb are built around: public structure, private depth.