A phData Cortex Agent surfaces natively in Snowflake CoWork — but as an agent object, not a "Skill"; the cross-tenant partner publish path is still hardening
The question
"Can a phData-built Cortex Agent be natively published as a CoWork Skill so business users discover it inside the CoWork UX without a custom app, and what is the Snowflake partner publish path?" Context: if phData deliverables surface through CoWork's native store, the pitch becomes "the client's users get it for free through CoWork" and time-to-adoption drops; if not, every phData agent needs a bespoke Streamlit or MCP app — which changes how the DSA scopes and prices delivery. Open follow-up #4 from [[2026-07-07-snowflake-intelligence-vs-cortex-ai-boundary]]. ("CoWork" = Snowflake's Summit-2026 rebrand of Snowflake Intelligence — confirmed current name.)
What we already know (from the vault)
- Snowflake Intelligence (now Snowflake CoWork) is the user-facing agentic app layer; Cortex Agents are the orchestration objects underneath. The parent brief already concluded "a phData-built agent can then be surfaced through CoWork for business users, or through a custom Streamlit/MCP app — same Cortex plumbing, different face" — this brief pins down which. See [[2026-07-07-snowflake-intelligence-vs-cortex-ai-boundary]].
- phData's delivery spine builds at the Cortex Agents layer (
setup → create-views → create-semantic-view → create-agent → deploy, wrapped with golden-set / TruLens eval IP). The deliverable is a governed Cortex Agent, with CoWork as the optional business-user face. See [[2026-05-20-phdata-cortex-agents-practice]] and [[2026-07-07-snowflake-intelligence-vs-cortex-ai-boundary]]. - NAMING TRAP flagged in the vault: the CAF architecture doc's "Cowork" is Anthropic's Claude Cowork — the internal surface where phData's ~60 salespeople work, with org-admin-published Claude Skills — NOT Snowflake CoWork. This question is about Snowflake CoWork (the client's business-user surface). Two different products, both with a "skills" concept and org-admin publishing; do not conflate. See [[2026-06-15-caf-architecture-decisions-and-meta-council]].
- A "skill" is already a known unit in the Snowflake/Cortex Code world: SKILL.md capability packages (Claude Code-compatible), distributable org-wide via a Snowflake connection profile. The Brigade ships as exactly this. See [[2026-07-05-brigade-skills-snowflake-cortex-code]] and [[2026-07-03-hex-deployment-matrix]].
What the web says
- A Cortex Agent is a reusable account object (bundling model + tools + orchestration instructions), created in Snowsight, via Cortex Agents SQL, or the REST API. Business users "interact with them in Snowflake CoWork and Cortex Code" in addition to REST-API embedding — i.e. native surfacing is a first-class path, not app-only. (Cortex Agents docs)
- The native visibility mechanism is the "Snowflake Intelligence Object" — a single account-level container holding a curated list of available agents. Admins with ALTER/MODIFY privilege add/remove agents to control which appear to users; the Agent UI in Snowsight handles identity, access control, and monitoring out of the box. So a phData-built agent surfaces to business users with zero custom Streamlit/MCP app — you register the agent, add it to the Intelligence Object, grant RBAC. (Create and manage agents docs)
- "Skill" is a different unit than "agent." A skill is "a modular, portable package of instructions, scripts, and context that gives agents the capability to perform specific, repeatable tasks" — defined by a
SKILL.mdfile, stored in named Snowflake stages or Git repos, auto-discovered and invoked by agents during orchestration. Skills are capabilities agents use, not a wrapper you publish an agent as. (Agent skills docs) - CoWork adds a discovery surface for skills: User Skills + a Skill Catalog. "Teams can automate their unique workflows with User Skills and leverage the new Skill Catalog to easily discover, share, and reuse them"; users can also "discover and install shared Cortex Code skills and plugins directly from Snowflake CoWork." Skill Catalog and Agent Studio were flagged "coming to GA soon" at Summit (preview, not fully GA). (Snowflake CoWork press release, Atlan Summit 2026 recap)
- The emerging cross-enterprise publish/discover path is the Agentic Resource Discovery (ARD) Specification — an open protocol (Snowflake + Microsoft + GoDaddy) with a Describe→Curate→Search→Execute flow: publishers post an
ai-catalog.jsonmanifest; discovery services index it; clients search in natural language and connect via MCP/A2A/REST. Snowflake states that when a Cortex Agent is published in Snowsight or via the SDK, "Snowflake could register the agent in the organization's discovery endpoint automatically," making it discoverable across CoWork. (ARD Specification blog) - Partner/productized distribution runs through the Marketplace track, and it is under-documented. Snowflake Marketplace + Native App Framework + Cortex Knowledge Extensions (CKE) are the partner surfaces; Snowflake is also a launch partner in Anthropic's Claude Marketplace. But the ARD blog and the skills/agent docs contain no explicit consulting-partner "publish an agent so another org's CoWork surfaces it" workflow yet. (Getting started with CoWork and CKE, ARD blog)
Convergences and contradictions
- Convergence — the "free UX" claim holds at the account level. Vault ("surfaced through CoWork") and web (native agent interaction in CoWork + the Snowflake Intelligence Object) agree: inside a single client account, a phData Cortex Agent reaches business users with no custom app. The Streamlit/MCP app becomes the exception, not the default.
- Sharpening — the question's premise mislabels the unit. You do not "publish an agent as a CoWork Skill." An agent surfaces as an agent (via the Snowflake Intelligence Object); a Skill is a separate SKILL.md capability package agents consume and the Skill Catalog surfaces. Both are legitimate phData packaging surfaces, but the free-UX benefit comes from the agent-object path, not from "publishing a skill."
- Contradiction/limit — "native store" ≠ cross-tenant app store. Native surfacing is per-account, done inside the client's own Snowflake. There is no confirmed "publish once, every client's CoWork users discover it" marketplace for agents today; that productized-reuse path (Marketplace / Native App / CKE / ARD) is real in direction but under-documented and partly preview.
Synthesis for RDCO
The headline for the DSA pitch is genuinely good, with one precise caveat. For a standard engagement, a phData-built Cortex Agent reaches the client's business users natively through Snowflake CoWork with no custom front-end build: create the agent object (Snowsight/SQL/REST), add it to the account's Snowflake Intelligence Object, and grant RBAC — CoWork renders the chat UX, identity, access control, and monitoring for free. That flips the old default. The bespoke Streamlit or MCP app moves from "every agent needs one" to "only when the client wants embedded/branded UX outside Snowflake, a non-CoWork portal, or a specialized interaction pattern." In scoping terms, this deletes a recurring front-end line item and re-weights the estimate toward where phData's actual value sits: the governed Cortex layer (semantic view for Analyst, Search services with RBAC-aware retrieval, agent configuration and tool boundaries, eval harness) plus the visibility/RBAC/Intelligence-Object wiring. Pricing narrative shifts from "we build you an app" to "we make the layer under CoWork trustworthy and governed, and CoWork gives your users the door for free."
But the DSA must not over-claim a marketplace that doesn't exist yet. "Natively surfaces in CoWork" is an in-account capability, not phData publishing an accelerator to a cross-client CoWork store that any customer installs. If the client hears "your users get it for free through CoWork," that is accurate for the agent phData stands up in their account. It is not accurate as "phData ships one agent and all our clients' users discover it." That cross-tenant, productized-IP motion is the Snowflake Marketplace / Native App Framework / Cortex Knowledge Extensions track — a different SKU, different validation, and partly preview (Agent Studio and Skill Catalog were "GA soon" at Summit; the ARD partner-publish workflow is not yet documented). Selling reusable accelerators as public CoWork installs is premature; selling a governed agent that lives in the client's CoWork is deliverable today.
Skills are the complementary packaging lever, and they connect to the Brigade thesis. Beyond the agent, phData can codify repeatable capabilities as SKILL.md skills (Claude Code-compatible per [[2026-07-05-brigade-skills-snowflake-cortex-code]]) stored in stages/Git, which the delivered agent auto-discovers and which the client's Skill Catalog/User Skills surface for reuse. This is where phData IP becomes reusable, governed, discoverable within the client account — and it dovetails with the Brigade already shipping as a native Cortex Code plugin distributable via a Snowflake connection profile. So the packaging story is two-tier: (1) the agent rides CoWork's native UX for end users; (2) skills package the repeatable capability IP the agent runs on, surfaced in the client's Skill Catalog.
Naming discipline for the field. In a client conversation, translate "publish it as a CoWork Skill" to "register the agent object and add it to the Snowflake Intelligence Object so it appears in CoWork; optionally package repeatable capabilities as Skills in the Skill Catalog." And keep the two Coworks apart: the client's Snowflake CoWork (where the delivered agent surfaces) versus phData's internal Anthropic Claude Cowork (where the sales org and CAF skills live). Blurring them will confuse both the client and the internal team.
Open follow-ups
- Pin the GA status at the client's contract date: Skill Catalog, Agent Studio, and User Skills were "GA soon" at Summit — which are GA vs preview in H2 2026, and does that gate whether phData can promise skill-based reuse?
- Is there a today-shippable cross-tenant path — publish a Cortex Agent (or agent + skills bundle) to Snowflake Marketplace / as a Native App so a different org's CoWork surfaces it — and what is the partner listing/validation/security-review process?
- How do Cortex Knowledge Extensions (CKE) differ from publishing an agent — is CKE strictly the partner path for reusable retrieval content/knowledge, and does an accelerator combine a CKE + an agent template?
- Does ARD's
ai-catalog.jsongive 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? - In a shared-services / managed engagement, can the Snowflake Intelligence Object curation model express phData-managed vs client-managed agent visibility (who controls which agents appear)?
Related
- [[2026-07-07-snowflake-intelligence-vs-cortex-ai-boundary]] — parent brief; establishes the CoWork (app layer) vs Cortex Agents (orchestration) split this brief resolves at follow-up #4
- [[2026-07-05-brigade-skills-snowflake-cortex-code]] — SKILL.md skills as native Cortex Code packages; the complementary skill-packaging lever
- [[2026-06-15-caf-architecture-decisions-and-meta-council]] — the Anthropic-Claude-Cowork vs Snowflake-CoWork naming trap; org-admin skill-distribution rails
- [[2026-05-20-phdata-cortex-agents-practice]] — phData's Cortex Agents delivery spine and deal shape
- [[2026-07-03-hex-deployment-matrix]] — packaging/runtime matrix; Cowork skills rail and deployment profiles
Sources
Vault:
- [[2026-07-07-snowflake-intelligence-vs-cortex-ai-boundary]] —
~/rdco-vault/06-reference/research/2026-07-07-snowflake-intelligence-vs-cortex-ai-boundary.md - [[2026-07-05-brigade-skills-snowflake-cortex-code]] —
~/rdco-vault/06-reference/research/2026-07-05-brigade-skills-snowflake-cortex-code.md - [[2026-06-15-caf-architecture-decisions-and-meta-council]] —
~/rdco-vault/01-projects/phdata/2026-06-15-caf-architecture-decisions-and-meta-council.md - [[2026-05-20-phdata-cortex-agents-practice]] —
~/rdco-vault/06-reference/research/2026-05-20-phdata-cortex-agents-practice.md - [[2026-07-03-hex-deployment-matrix]] —
~/rdco-vault/01-projects/phdata/2026-07-03-hex-deployment-matrix.md
Web:
- Cortex Agents — Snowflake docs (agent as reusable object; interacts in CoWork + Cortex Code): https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-agents
- Create and manage agents — Snowflake docs (Snowflake Intelligence Object; admin curation of agent visibility): https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-agents-manage
- Agent skills — Snowflake docs (skill = SKILL.md capability package agents consume; stages/Git storage): https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-agents-skills
- Overview of Snowflake CoWork — Snowflake docs: https://docs.snowflake.com/en/user-guide/snowflake-cortex/snowflake-cowork
- Snowflake CoWork press release (User Skills, Skill Catalog "GA soon"): https://www.snowflake.com/en/news/press-releases/snowflake-cowork-powers-the-agentic-enterprise-as-the-personal-agent-for-knowledge-workers-to-work-smarter/
- Agentic Resource Discovery Specification — Snowflake blog (ai-catalog.json; Describe/Curate/Search/Execute): https://www.snowflake.com/en/blog/agentic-resource-discovery-specification/
- Getting Started with Snowflake CoWork and Cortex Knowledge Extensions — Snowflake: https://www.snowflake.com/en/developers/guides/getting-started-with-cowork-and-cke/
- Atlan — Snowflake Summit 2026 announcements (Skill Catalog / Agent Studio status): https://atlan.com/know/snowflake/summit-2026-announcements/