Should Snowflake launch its own AI agent marketplace?
Yes, Snowflake should absolutely build an agent marketplace, but it should extend the existing Snowflake Marketplace + Native Apps Framework rather than spin up a separate "Agent Marketplace" sub-brand. Salesforce learned this lesson the hard way: AgentExchange launched as a standalone surface in 2025 and within months was being folded back into AppExchange because partners refused to maintain two listings, two billing flows, and two install paths.
Snowflake already has the rails - Marketplace for distribution, Native Apps for packaging, Snowpark Container Services for runtime, and Cortex Agents for the orchestration layer - so the strategic move is to add an agent listing type, not a new storefront. The internal pitch is simple: every agent is just a Native App that happens to have a planner loop.
Treat it that way and you ship in a quarter; treat it as a new product and you re-litigate every partner contract.
What Snowflake Already Has
- Snowflake Marketplace - live data and app distribution surface with thousands of listings, billing handled in Snowflake credits, install-into-account flow already familiar to admins.
- Native Apps Framework - packaging, versioning, security model, and consumer/provider separation already production-grade; agents are just apps with a
run_agent()entrypoint. - Snowpark Container Services - GA runtime for long-running stateful workloads, which is exactly what agents need (vs. Stateless UDFs).
- Cortex Agents - first-party agent primitives (Cortex Analyst, Cortex Search, tool-calling) that partner agents can compose against.
- Horizon governance - lineage, masking, access policies that already work across Marketplace installs; agents inherit this for free.
What's Missing For Agent Marketplace
- Agent-discovery taxonomy - no "Sales SDR agent" vs "Finance close agent" vs "DataOps agent" categorization in the current Marketplace UI.
- Cross-customer agent benchmarking - no public quality scores, no "this agent resolves 73% of tickets at p50" leaderboard equivalent.
- Agent-result attribution back to data - when an agent writes to a table or triggers a workflow, no first-class lineage that says "agent X did Y at time Z because of data W."
- Agent-pricing standardization - credits-per-token, credits-per-task, credits-per-resolution - partners are inventing one-off SKUs instead of picking from a menu.
- Human-in-the-loop framework - no shared approval/escalation UI; every agent partner is rebuilding the same Slack/email approval flow.
- Agent observability contract - no required telemetry schema for agent runs, so customers can't compare two competing SDR agents on the same axes.
Competitive Pressure
- Salesforce AgentExchange + Agentforce - launched March 2025, already has hundreds of partner agents and is the loudest voice in the category; the Data Cloud + Agentforce combo is the most direct threat to Snowflake's "AI Data Cloud" positioning.
- ServiceNow AI Agent Studio + Now Assist - shipping agents tied directly to workflow execution, which Snowflake structurally cannot match without a workflow engine.
- Databricks Mosaic AI Agent Framework - the closest architectural peer; if Databricks ships a marketplace surface first, the "data + agents in one place" story flips against Snowflake.
- Anthropic Apps + OpenAI GPT Store / Agent Builder - foundation-model-vendor marketplaces that don't care about your data warehouse and will happily disintermediate Snowflake for the chat layer.
- Microsoft Copilot Studio agent gallery - distribution via Microsoft 365 is a category killer for any agent that touches Office data; Snowflake needs a counter for the data-side workloads.
The Build Plan
- Add an
agentlisting type to Marketplace - same install flow, new badge, new filter facet. Two-sprint UI change. - Standardize agent pricing in credits - publish three SKU templates (per-token, per-task, per-resolution) and require partners to pick one. Kills bespoke billing negotiations.
- Launch with named partners - Hightouch (reverse-ETL agent), Snowflake Cybersyn (research agent), Salesforce Data Cloud connector (cross-CRM agent), and the Cortex Cookbook reference agents. Five logos at GA, not fifty.
- Agent-quality scoring - publish a benchmark harness (eval set + scoring rubric) and show scores on each listing. Steal the model-card pattern from Hugging Face.
- Customer attribution dashboard - one Snowsight page per installed agent: runs, tokens, credits, business outcome metric the agent declared. Turns agents into measurable line items.
- Native HITL primitive - ship a
cortex.request_approval()function with a default Slack/email/Snowsight inbox. Removes the #1 partner re-build cost. - Free Cortex Agents tier for first 90 days post-install - subsidize trial, capture the data on which agents convert.
Why Sub-Brand Would Backfire
- Sales-team confusion - Snowflake AEs already sell Marketplace, Native Apps, Cortex, Horizon, and Polaris; a sixth surface area is one too many for a quarterly QBR slide.
- Partner economics fragmentation - partners would have to negotiate separate rev-share, separate listing fees, separate certification for the Agent Marketplace vs. Marketplace; most will pick one and skip the other.
- Dilutes Marketplace traffic - the Marketplace SEO and in-product discovery surface is a hard-won asset; splitting it halves the funnel for everyone.
- Failed precedents - Salesforce AgentExchange was quietly re-merged with AppExchange after the standalone storefront underperformed; ServiceNow's earlier Service Graph Connector marketplace also got absorbed back into the main Store.
- Re-platforming cost on the customer side - admins do not want a second install/governance/billing surface to learn; the friction shows up immediately in adoption metrics.
Strategy Options
| Strategy | Capex | Time-to-launch | Partner adoption risk | Recommendation |
|---|---|---|---|---|
| Standalone Agent Marketplace (sub-brand) | High | 9-12 months | High (two-listing tax) | Avoid |
| Extend existing Snowflake Marketplace | Medium | 3-4 months | Low (one listing, one billing) | Recommended |
| Cortex-only agent gallery (first-party only) | Low | 6 weeks | High (closes out partners) | Avoid - cedes ecosystem |
| Acquire and rebrand a third-party agent hub | Very high | 6-9 months + integration | Medium (channel conflict) | Only if a category-defining target appears |
| Do nothing | Zero | n/a | Catastrophic vs. Salesforce + Databricks | Not viable |
Strategic Flow
FAQ
Should Snowflake build a separate Agent Marketplace or extend the existing one? It should extend the existing Snowflake Marketplace plus Native Apps Framework rather than spin up a separate "Agent Marketplace" sub-brand. The recommended move is to add an agent listing type, treating every agent as just a Native App that happens to have a planner loop, which ships in a quarter instead of re-litigating every partner contract.
What Salesforce precedent warns against a standalone storefront? Salesforce AgentExchange launched as a standalone surface in 2025 and within months was being folded back into AppExchange because partners refused to maintain two listings, two billing flows, and two install paths.
ServiceNow's earlier Service Graph Connector marketplace was also absorbed back into the main Store.
What rails does Snowflake already have for hosting agents? Snowflake already has the Marketplace for distribution with billing in credits, the Native Apps Framework for packaging and versioning, Snowpark Container Services as the GA runtime for long-running stateful workloads, and Cortex Agents as first-party orchestration primitives.
Horizon governance gives agents lineage, masking, and access policies for free.
What is missing before an agent marketplace can launch? Gaps include an agent-discovery taxonomy (no "Sales SDR agent" vs "Finance close agent" categories), cross-customer agent benchmarking and quality leaderboards, agent-result attribution back to data, standardized agent pricing, a shared human-in-the-loop approval framework, and a required agent observability telemetry schema.
Which named partners should launch the agent listing type? The build plan calls for launching with five logos at GA, not fifty: Hightouch (reverse-ETL agent), Snowflake Cybersyn (research agent), the Salesforce Data Cloud connector (cross-CRM agent), and the Cortex Cookbook reference agents.
It also recommends a cortex.request_approval() HITL primitive and a free Cortex Agents tier for the first 90 days post-install.
Bottom Line
Ship the agent marketplace, but ship it as a listing type inside the existing Snowflake Marketplace, not as a standalone storefront. The competitive window vs. Salesforce Agentforce and Databricks Mosaic is measured in quarters, not years; a sub-brand burns two of those quarters on internal re-platforming for zero customer benefit.
Use the Native Apps Framework as the substrate, Cortex Agents as the first-party reference, and five named partners as the GA proof. Win on time-to-credit-revenue, not on storefront aesthetics. (see also: q1564, q1583, q1595, q1600)