The Machines Have a Directory Now.
People Are Still Waiting for Theirs.

In June, Google announced the specification for how AI agents will find everything they use — publisher-controlled catalogs, federated registries the announcement calls “search engines for the agentic web.” Its current file carries authors from Google, Microsoft, and Hugging Face, and every entry it can return describes something an agent can call. This is not a story about what the spec got wrong. It is a reading of what got formalized first.

By William Douglas Baker

Founding Editor

On June 17, 2026, on the Google Developers Blog, two engineers announced the layer every other agentic announcement had been quietly assuming: Agentic Resource Discovery — “An open specification for finding and verifying tools, skills, and agents across the web.” The design is two primitives. An organization publishes a catalog of its entries at a well-known path on its own domain; the current file names that path /.well-known/ard.json, keeping the announcement’s ai-catalog.json as a predecessor courtesy. Federated registries crawl those catalogs, index them, and answer plain-language discovery queries; the announcement calls registries “search engines for the agentic web,” and means it as a job description, not a metaphor. The announcement says the spec was “Developed with partners across the industry” and builds on a Linux Foundation working group’s data model; the specification itself stands at version 0.91, status Proposal, current file dated August 26, 2026, its author list drawn from Google, Microsoft, and Hugging Face. It was re-pulled and re-read in full for this edition at press time. It is careful, unglamorous, load-bearing work: the plumbing decision that determines what an agent can find at all.

The announcement opens with the three questions the layer exists to answer: “Where does the right capability live? Which capability should I actually use? And how do I verify it’s safe to connect to?” The spec’s answer is an entry format of unusual discipline. An entry describes “a single agentic resource in a form that can be found by search.” Its type field is mandatory and takes an IANA media type: an MCP server card, an agent card, a skill. Its scope is declared in the opening paragraph: “MCP tools, A2A agents, skills, and other callable services.” Trust is anchored where the web can actually verify it — the announcement: “ownership of that domain serves as the cryptographic foundation for identity and trust.” In the spec’s own terms, the discovery identifier is “a stable searchable handle,” and the security principal beneath it is “a dynamic cryptographic credential.” When an entry supplies trust metadata, registries are expected to verify it under the declared framework and may use the result in filtering, ranking, or admission; semantic relevance itself remains a separate score. Whoever wrote this has watched a namespace get squatted.

Now read it for what it does not contain, as an inventory rather than an accusation. Across the current specification there is no entry type for a person and no field for professional licensure. The words person and individual never appear. The one credential the document defines belongs to the machine — the cryptographic kind, not the professional kind. Identity is domain-anchored: when an entry carries cryptographic trust metadata, the asserted trust domain must align with the publisher domain in its discovery identifier. And none of this is a defect, because none of it is the document’s job. A discovery layer for callable resources owes nobody a category for human judgment; judgment is not callable. Prosecuting this spec for lacking a person type would be reviewing a bridge for its bedside manner. The finding is quieter, and larger: in this specification, the industry has said precisely and in committee how a weather tool should be found. The same file does not define how a licensed professional should be.

The author list carries the detail that turns an observation into a chronology. Among the spec’s three named authors is Microsoft’s R.V. Guha — the engineer schema.org’s own documentation says “initiated schema.org and is one of its co-founders.” Schema.org is the vocabulary that made the human web machine-readable, and it carries a Person type defined with room to spare: “A person (alive, dead, undead, or fictional).” Schema.org gave the human web a Person type. ARD v0.91 gives the agentic web an extensible entry for callable resources. Guha appears in the author history of both. The chronology does not prove motive. It does show what each generation chose to formalize first.

What got formalized first

This record spent yesterday’s edition on the answer layer’s arithmetic: a randomized experiment showing the interface can keep the click without delivering the compensating experience. Edition No. 055 documented the extraction; this one documents the architecture rising behind it. Three generations, one direction. The ten blue links showed a person the candidates and let them choose. The answer layer chooses first and shows its citations. The agentic layer inserts a registry between the question and the candidates — and its entry format types what may be found before any ranking begins. In the clinical vocabulary of this record, that is a structural precondition for Decision Exclusion: not a decision made against an entity, but a candidate pool whose native schema does not represent the entity as a person. The condition this record documents daily does not begin at the ranking. It begins at the catalog.

The assignment, then, is not to petition the agentic web for a person slot; standards bodies move at their own tempo, and this one is honest about its scope. The assignment is to notice what the sequence makes scarce. The cited record supplies a written standard for machine discoverability. It supplies no comparable specification for how a name, a license, and a body of judgment become legible to systems answering on a buyer’s behalf. The human side has practice: the work this record’s publisher calls Answer Engine Authority™, encoding recognition where engines already look while the registries are still deciding what to read. In this record, the corresponding human-authority layer remains unwritten. Unwritten layers belong to whoever writes them first.