No. "ARD" Is Agentic Resource Discovery, Not Agent Ready Data; It Is a v0.91 Proposal Whose ai-catalog.json Filename Is Already Deprecated, and docs.snowflake.com Documents None of It

The question

"Does Agent Ready Data (ARD) ai-catalog.json give phData a supported way to register accelerator agents for enterprise-wide auto-discovery inside a client's CoWork, and is that mechanism GA or spec-only?" This is open follow-up #4 from [[2026-07-08-cortex-agent-cowork-skill-native-publish]], auto-promoted 2026-07-22 at 13/15.

Headline: The premise carries a name error and the answer is a clean negative on three independent grounds. (1) ARD stands for Agentic Resource Discovery, not "Agent Ready Data"; the two are unrelated ideas that the vault has been tracking separately. (2) The spec is v0.91, "Status: Proposal", dated 2026-08-26, and as of that revision ai-catalog.json is explicitly the predecessor path, superseded by /.well-known/ard.json. (3) Snowflake's only public statement is a 2026-06-17 marketing blog written in the conditional ("Snowflake could register the agent"), and a direct sweep of docs.snowflake.com found zero occurrences of ARD, ai-catalog, "Agentic Resource Discovery", "well-known" or "discovery endpoint" on the CoWork, Cortex Agents, agent-management and agent-sharing pages, in the 8,743-URL documentation sitemap, or in the all-release-notes index. Verdict: spec-only, and not even the current spelling of the spec.

What we already know (from the vault)

• The parent brief already flagged ARD as the emerging cross-enterprise publish path and quoted Snowflake's exact conditional: agents "could" be auto-registered in the organization's discovery endpoint. It also recorded that the ARD blog and the agent docs contain "no explicit consulting-partner 'publish an agent so another org's CoWork surfaces it' workflow." See [[2026-07-08-cortex-agent-cowork-skill-native-publish]].
• That gap has since been closed by a different mechanism. The GA cross-tenant path is the Snowflake Native App carrying a Cortex Agent (release note 2026-08-07), with standalone agent sharing still open preview and unable to carry skills, procedures or MCP connector tools. ARD is not part of either path. See [[2026-09-23-cortex-agent-cross-tenant-marketplace-publish]].
• The house evidentiary rule is already written down: "the release note is the evidence; the press release is not." A feature may be promised as production-grade only if a docs.snowflake.com release note says General Availability and the feature doc carries no preview banner. ARD fails this test at the first hurdle, having no docs page at all. See [[2026-09-14-snowflake-cowork-cortex-ga-vs-preview-matrix]].
• "Agent-Ready Data" is a separate vault thread, and it is an architecture argument, not a product. The vault tracks it via the Sadalage and Chandrasekaran piece (five attributes: Trusted, Contextual, Traceable, Governed, Operational) about making data legible to agents. It has no manifest file and no registration semantics. See [[2026-08-31-dataengineeringweekly-285-agent-ready-data-architecture]].
• The vault has already concluded that vendor control planes have no incentive to cede agent discovery to a neutral layer, and that cross-vendor agent state has no standard even where identity is converging. ARD's slow adoption fits that finding rather than contradicting it. See [[2026-09-13-cross-substrate-agent-identity-registry-standard]].

What the web says

• ARD = Agentic Resource Discovery. The canonical repo ards-project/ard-spec describes itself as "a federated, domain-anchored standard for cataloging, searching, and discovering agentic resources (MCP servers, A2A agent cards, Skills, APIs, and other callable services)". Apache-2.0, created 2026-05-19, last pushed 2026-09-12, 468 stars, 59 open issues. Nothing in it uses the phrase "Agent Ready Data". (ard-spec repo, read directly plus GitHub API metadata)
• Status is a proposal, not a ratified standard. The spec header reads verbatim: **Version**: v0.91, **Status**: Proposal, **Date**: August 26, 2026. Authors are Junjie Bu (Google), R.V. Guha, and Shaun Smith (Hugging Face). There is no standards body, no consortium charter, and no conformance requirement on any vendor. (spec/ard.md)
• ai-catalog.json is deprecated inside the spec itself. v0.91, normative: "A consumer resolving a domain's entries MUST fetch /.well-known/ard.json ... ARD's predecessor specified the path /.well-known/ai-catalog.json and the link relation ai-catalog; a consumer MAY additionally consult these". Informative follow-on: "A resource that remains only at /.well-known/ai-catalog.json may not be found, since consulting that path is optional for consumers - a publisher on the predecessor path SHOULD move to ard.json." Building to the filename in the question would be building to a legacy name. (spec/ard.md §5.1)
• Snowflake's position is an endorsement blog, in the conditional, three months stale. The 2026-06-17 post names Microsoft, GoDaddy and "many others" as co-developers and describes the vision: "On publish, Snowflake could register the agent in the organization's discovery endpoint automatically." No GA, public preview or private preview status is stated anywhere in it. Source type: vendor marketing blog, weak evidence by our own standard. (Snowflake ARD blog)
• Primary-docs sweep returns a clean zero. Fetched and text-searched four docs.snowflake.com pages in full (CoWork overview, Cortex Agents, Create and manage agents, Share Cortex Agents): no hit on ai-catalog, "Agentic Resource Discovery", "well-known", "discovery endpoint" or "registry". The 8,743-URL docs sitemap contains no slug matching ard, agentic or discovery (the only near-match is the Google Ads gard connector). The all-release-notes index likewise returns nothing. The only preview banner found anywhere in the sweep was the one already known: "Preview Feature - Open" on Share Cortex Agents. (Cortex Agents, Create and manage agents, Share Cortex Agents, CoWork overview)
• Live adoption probe: thin, and mixed even among the authors' own employers. Only huggingface.co/.well-known/ai-catalog.json returned HTTP 200 with a real manifest (a Hugging Face Discover Registry entry, specVersion 1.0); its ard.json returned 401. google.com returned 404 on both paths. phdata.io returned 404 on both, so phData publishes no manifest today. snowflake.com, microsoft.com and godaddy.com returned 403 - bot/WAF blocks, so those three are inconclusive, not negative, and are flagged rather than counted either way.
• Architecturally, ARD is a public-web mechanism, which is the deeper mismatch. Manifests are domain-anchored and served over HTTPS from a publisher's own domain; discovery services crawl them; enterprises then curate what enters their registry; clients search in natural language and connect directly, bypassing the registry for invocation. There is a deliberate human curation step in the middle, so even a fully shipped ARD would not be "auto-discovery" into a client tenant without that client's opt-in. (spec/ard.md §3, §5)

Convergences and contradictions

• Convergence, and the parent brief holds up well. On 2026-07-08 the vault called ARD "real in direction but under-documented"; ten weeks later it is still exactly that, while the adjacent Native App path went GA in the same window. The parent's instinct to not build the pitch on ARD was correct.
• Contradiction inside the question itself. The question asks whether ai-catalog.json gives a supported registration path. As of the current spec revision the file is the old name, so the honest answer is not just "not GA" but "not the current artifact either". A brief that answered only the GA half would leave a trap in place for anyone who later greps the vault for the filename.
• Contradiction between announcement and documentation. Snowflake's marketing surface says it supports ARD; Snowflake's documentation surface says nothing at all. Under the vault's own release-note rule, the documentation surface wins and the capability does not exist for contract purposes. This is the exact press-release-versus-release-note gap that [[2026-09-14-snowflake-cowork-cortex-ga-vs-preview-matrix]] was written to police.
• A weaker signal worth naming honestly. Hugging Face serving a live manifest proves the format is more than a slide, and 468 stars plus nine architecture decision records show real engineering activity. That is evidence of a healthy early spec, not evidence of a shippable Snowflake feature. The two should not be blurred.

Synthesis for RDCO

The operative answer for the DSA is one sentence: there is no ARD-based registration path to rely on, and the thing to rely on instead already exists. If a phData accelerator agent needs to reach users in a client's CoWork, the GA mechanism is the one [[2026-09-23-cortex-agent-cross-tenant-marketplace-publish]] pinned down: package it as a Snowflake Native App whose setup script runs CREATE AGENT, distribute through a Marketplace or private listing, and the agent appears in that client's CoWork. In-account, the mechanism is the Snowflake Intelligence Object plus role-based access control. Both are documented, both have release notes, and neither depends on a proposal-stage web standard. ARD is a roadmap line, not a delivery line, and it should never appear in a Statement of Work except under "subject to Snowflake availability".

The name collision is the finding most likely to cause real damage, and it should be corrected on contact. "Agent Ready Data" and "Agentic Resource Discovery" are different enough that a client or an internal reviewer who hears "ARD" will reasonably guess wrong. The first is a data-architecture posture about trust, context and traceability in the data layer, which is phData's actual selling ground and maps directly to the semantic-view and governance work in [[2026-05-20-phdata-cortex-agents-practice]]. The second is a discovery protocol about where a manifest of agent descriptions lives on the public internet. phData can credibly sell the first today with no vendor dependency. It cannot sell the second at all. Anyone who collapses the two produces a pitch that sounds like a shipped integration and is in fact a JSON file nobody parses on the client's behalf.

Even the mature version of ARD would not deliver the "enterprise-wide auto-discovery" the question imagines, and that is worth internalizing now rather than re-litigating in 2027. The spec's own flow puts a curation step between crawling and availability: a discovery service indexes manifests, then the enterprise decides what enters its registry. That is a governance gate by design, and it is the same shape as the consumer-side consent steps already documented on the Native App and agent-sharing paths (the "Add to Snowflake CoWork" toggle, the GRANT CALLER grants under restricted caller's rights). No enterprise buying Snowflake is going to let an outside vendor's manifest inject agents into its employees' assistant without a human approving it. So the realistic ceiling for ARD, whenever it ships, is "phData's accelerator is findable and one approval away", not "it appears automatically". That is a marketing-reach improvement, not a distribution breakthrough, and it should be priced as such.

The watch condition is cheap and specific, so set it and stop re-researching. ARD becomes worth revisiting only when a docs.snowflake.com page or release note names it, at which point the GA-versus-preview matrix gets a new row. Until then the correct posture is the one this brief recommends: quote the Native App path, describe agent-ready data work in phData's own terms, and treat ARD as a standards-watch item. If phData ever does want to participate cheaply, publishing a /.well-known/ard.json on phdata.io (note: ard.json, not ai-catalog.json, and phdata.io currently serves neither) is a low-cost marketing experiment for public-web agent discovery. It is unrelated to anything happening inside a client tenant and should not be confused with one.

Why this is in the vault

It settles whether phData can rely on an auto-discovery registration path for accelerator agents in client CoWork tenants: it cannot, so accelerator packaging decisions and any Statement of Work language stay anchored to the GA Native App path from [[2026-09-23-cortex-agent-cross-tenant-marketplace-publish]] rather than to the ARD roadmap. It also fixes a name collision ("Agent Ready Data" versus "Agentic Resource Discovery") that would otherwise turn a proposal-stage web spec into an implied shipped integration in a client conversation.

Open follow-ups

• Does Snowflake's snowflake.com actually serve an ARD manifest? The 403 from the bot filter left this genuinely unresolved, and a browser-based check (Playwright, per the vault's Cloudflare-block pattern) would settle whether the endorsement is backed by a live artifact.
• Does the Snowflake Horizon Catalog, which the docs describe as "the agentic catalog for all your data, whether it is inside or outside of Snowflake", already implement an internal agent-registry function that ARD would later federate into? If so, Horizon is the surface to watch, not ARD.
• Has ARD moved beyond Status: Proposal since 2026-08-26, and has any vendor shipped a conformant discovery service rather than only a manifest? The conformance tooling in the repo would make this objectively checkable.
• If a client asks for cross-vendor agent discovery today (Snowflake agents plus Microsoft plus internal), what is the actual buildable answer, and does it reduce to a per-vendor registry plus a thin aggregator that phData would have to own and maintain?
• Is there a defensible phData service offer in the Agent-Ready Data sense (the five-attribute data-layer posture) that can be positioned without any dependency on vendor discovery standards, and what would its assessment artifact look like?

Related

• [[2026-07-08-cortex-agent-cowork-skill-native-publish]] - parent brief; this closes its open follow-up #4 and confirms its cautious read of ARD
• [[2026-09-23-cortex-agent-cross-tenant-marketplace-publish]] - the GA cross-tenant path that ARD does not replace and should not be confused with
• [[2026-09-14-snowflake-cowork-cortex-ga-vs-preview-matrix]] - the release-note-is-the-evidence rule that decides this brief's verdict
• [[2026-07-07-snowflake-intelligence-vs-cortex-ai-boundary]] - the CoWork-versus-Cortex-Agents layer split the discovery question sits on top of
• [[2026-05-20-phdata-cortex-agents-practice]] - phData's delivery spine; where agent-ready data work actually lands
• [[2026-08-31-dataengineeringweekly-285-agent-ready-data-architecture]] - the separate "Agent-Ready Data" architecture thread the question's name collided with
• [[2026-09-13-cross-substrate-agent-identity-registry-standard]] - prior vault finding on why cross-vendor agent registries stall
• [[2026-07-25-portable-agent-memory-standard]] - companion standards-watch brief on portability gaps between agent substrates

Sources

Vault:

• [[2026-07-08-cortex-agent-cowork-skill-native-publish]] - ~/rdco-vault/06-reference/research/2026-07-08-cortex-agent-cowork-skill-native-publish.md
• [[2026-09-23-cortex-agent-cross-tenant-marketplace-publish]] - ~/rdco-vault/06-reference/research/2026-09-23-cortex-agent-cross-tenant-marketplace-publish.md
• [[2026-09-14-snowflake-cowork-cortex-ga-vs-preview-matrix]] - ~/rdco-vault/06-reference/research/2026-09-14-snowflake-cowork-cortex-ga-vs-preview-matrix.md
• [[2026-07-07-snowflake-intelligence-vs-cortex-ai-boundary]] - ~/rdco-vault/06-reference/research/2026-07-07-snowflake-intelligence-vs-cortex-ai-boundary.md
• [[2026-05-20-phdata-cortex-agents-practice]] - ~/rdco-vault/06-reference/research/2026-05-20-phdata-cortex-agents-practice.md
• [[2026-08-31-dataengineeringweekly-285-agent-ready-data-architecture]] - ~/rdco-vault/06-reference/2026-08-31-dataengineeringweekly-285-agent-ready-data-architecture.md
• [[2026-09-13-cross-substrate-agent-identity-registry-standard]] - ~/rdco-vault/06-reference/research/2026-09-13-cross-substrate-agent-identity-registry-standard.md
• [[2026-07-25-portable-agent-memory-standard]] - ~/rdco-vault/06-reference/research/2026-07-25-portable-agent-memory-standard.md

Web (primary):

• ARD specification text, v0.91, Status: Proposal, 2026-08-26 - https://raw.githubusercontent.com/ards-project/ard-spec/main/spec/ard.md
• ARD spec repository and metadata (Apache-2.0, created 2026-05-19, 468 stars, 59 open issues) - https://github.com/ards-project/ard-spec
• Snowflake docs, Cortex Agents - https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-agents
• Snowflake docs, Create and manage agents - https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-agents-manage
• Snowflake docs, Share Cortex Agents (Preview Feature - Open) - https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-agents-sharing
• Snowflake docs, Overview of Snowflake CoWork - https://docs.snowflake.com/en/user-guide/snowflake-cortex/snowflake-cowork
• Snowflake docs sitemap, 8,743 URLs, swept for ARD slugs - https://docs.snowflake.com/sitemap.xml
• Snowflake docs, all release notes index, swept for ARD terms - https://docs.snowflake.com/en/release-notes/all-release-notes
• Live manifest probe, Hugging Face, HTTP 200 - https://huggingface.co/.well-known/ai-catalog.json

Web (vendor marketing, weak evidence, labeled as such):

• Snowflake blog, "Snowflake and the Agentic Resource Discovery Specification", 2026-06-17, conditional language throughout - https://www.snowflake.com/en/blog/agentic-resource-discovery-specification/

Blocked / inconclusive (flagged, not retried):

• snowflake.com, microsoft.com, godaddy.com /.well-known/ard.json and /.well-known/ai-catalog.json returned HTTP 403 from bot filtering. Treated as unknown, not as absence.