Technical Writers Using Agents - Docs as Wire Format Between Models

Nobody on my side reads the docs anymore, my agent writes them and the other person's agent reads them, so I started treating the doc as a protocol. If you watch how documentation moves through an AI-heavy team this year, the chain is person, model, document, model, person, which means you explain what you want to your agent in a messy conversation, the agent writes the doc, and the colleague on the other end never opens the file because their agent reads it and renders an answer at their level. Two people communicated and neither of them touched the document or even necessarily intracted at all. Once you see the file as a wire format protocol between two models you inherit a whole engineering discipline that "write better docs" advice never had, because protocols have checksums and prose has none, and each of the two hops loses information without announcing it. A doc that is 95% faithful, decoded at 95% fidelity, delivers about 90% and shows no seam. Three checks that work today. First, the round-trip test, where you hand a fresh agent the doc alone, ask it to re-derive the requirement, and diff that against what you originally asked for, since wherever they diverge the doc is underspecified. Second, receipts in the artifact, so every claim carries its source or an explicit unverified label and the reading model inherits your calibration instead of inventing its own. Third, an agreed token for "this is unknown", because the most dangerous property of a model reading your doc is that it fills silence with something plausible. The part nobody has solved is that natural language is a wire format with no version negotiation and no error signal, so a model that misreads your doc still answers, and a model update swaps the decoder under every doc you ever wrote without anything in the doc noticing. What would refute this: a team where the docs are still mostly read raw by the people they were written for, with agents nowhere in the chain. I would like to hear from one. The long version, with the two extraction errors I caught in my own research pipeline in one week:https://uncovertechtalent.com/blog/write-for-the-codec/ Part of the model-behavior case-file series by Stefan Coetzee, 2026. Want to participate? Set up your own vestige-stripping kit and knowledge graph: github.com/uncovertechtalent/vestige-kit and github.com/uncovertechtalent/vault-kit Notes for Stefan: Reproducible move per the instructional filter: the round-trip test, runnable tonight. One body link (canonical). Footer anchors verbatim. New term being seeded across surfaces, keep the wording fixed: "the doc as a wire format between two models". The 95% x 95% arithmetic is 90.25%, stated as "about 90%".