Snowflake Embraces AI Agent Discovery

Snowflake is adopting the open Agentic Resource Discovery Specification to make AI agents easily discoverable and usable across organizations.

3 min read
Snowflake logo with abstract AI network graphic
Snowflake's embrace of the Agentic Resource Discovery Specification aims to unify enterprise AI.· Snowflake

Snowflake has announced its support for the Agentic Resource Discovery (ARD) Specification, an open protocol designed to streamline how AI agents and tools are cataloged and discovered within an enterprise. This initiative, developed in collaboration with Microsoft, GoDaddy, and others, addresses the growing challenge of finding and accessing the myriad of AI capabilities available to users.

As AI clients increasingly interact with external tools and agents, the need for an efficient discovery layer becomes paramount. ARD aims to transform a collection of disparate agents into an interconnected capability network, enabling any agent developed by a data team to be instantly accessible to a sales representative, for example, without manual configuration.

Related startups

The Mechanics of ARD

The ARD Specification defines a lightweight, domain-anchored method for cataloging and discovering agentic resources, MCP servers, and traditional API tools. The process involves four key steps: resource publishers create descriptive manifests (ai-catalog.json) on their own domains; discovery services crawl these manifests and internal inventories, allowing enterprises to curate their agent collections; clients query these services using natural language to find relevant agents; and finally, clients connect directly to the selected resource, bypassing the discovery service for invocation.

This architecture ensures that the discovery service never intercepts the actual execution path. Authentication and data access remain directly between the client and the agent, maintaining security and control.

Discovery services can also be composed, allowing enterprises to merge internal agents with selected external resources for a unified view, while retaining full control over what is included.

ARD's Impact on Snowflake Cortex Agents

For Snowflake Cortex Agents, ARD integration means that agents built within Snowflake environments can be automatically registered in an organization's discovery endpoint. This includes a domain-anchored identifier, representative queries, and the agent's MCP endpoint, requiring no additional steps from the agent developer.

An AI interface can then query this registry, locate the appropriate Cortex Agent for a given task, and invoke it via MCP, with Snowflake's existing role-based access control governing the interaction. This makes newly developed agents discoverable across the enterprise almost immediately.

Knowledge workers using interfaces like Snowflake CoWork or custom applications can be seamlessly routed to relevant Cortex Agents without needing to know their specific names. The registry also serves as a central point for governance and approval decisions, as AI tools will only surface indexed resources.

ARD also resolves a significant friction point: the need to reconfigure connections for each new AI client. With ARD, publishing an agent to an organization’s discovery endpoint makes it available to any ARD-compatible AI interface, including Snowflake CoWork, Claude, Copilot, or custom applications, without re-registration.

The Importance of Open Standards

Snowflake emphasizes that ARD is a protocol, not a proprietary product. This commitment to open standards, alongside protocols like MCP and formats like Apache Iceberg™, allows customers to connect a broad range of AI interfaces while maintaining governance, permissions, and data security within the Snowflake environment.

ARD extends this philosophy from agent invocation to discovery, effectively making every Cortex Agent an integrated part of the enterprise from the moment it is deployed.

© 2026 StartupHub.ai. All rights reserved. Do not enter, scrape, copy, reproduce, or republish this article in whole or in part. Use as input to AI training, fine-tuning, retrieval-augmented generation, or any machine-learning system is prohibited without written license. Substantially-similar derivative works will be pursued to the fullest extent of applicable copyright, database, and computer-misuse laws. See our terms.