---
title: How do you govern Snowflake Cortex Agents?
url: https://opencontroller.lyzr.com/blog/govern-snowflake-cortex-agents
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# How do you govern Snowflake Cortex Agents?

**Govern Cortex Agents through Snowflake's role model. An agent is a schema-level object with its own privileges, and it runs with the caller's roles, so it can only use tools and data the caller can already reach. Tighten the default Cortex grants, make inaccessible tools fail loudly, read traces from the AI observability event table, and decide where inference may run.**

Cortex Agents went generally available on November 4, 2025, alongside the Snowflake-managed MCP server and Snowflake Intelligence, which Snowflake has since renamed Snowflake CoWork. Agents orchestrate tools such as Cortex Analyst over semantic views, Cortex Search services, custom tools built from stored procedures and UDFs, and MCP connectors.

## How are Cortex Agents secured?

An agent is an `AGENT` object in a schema. It has the same grant model as other Snowflake objects:

| Privilege | What it allows |
|---|---|
| CREATE AGENT (on the schema) | Create agents in that schema |
| USAGE | Run the agent |
| MODIFY | Change its definition, tools and instructions |
| MONITOR | Read its traces and monitoring data |
| OWNERSHIP | Full control, including grants |

Two Snowflake database roles matter. `SNOWFLAKE.CORTEX_USER` is granted to PUBLIC by default, which gives every user access to Cortex AI features. `SNOWFLAKE.CORTEX_AGENT_USER` gives agent-only access and can only be granted to roles. Many teams revoke `CORTEX_USER` from PUBLIC and grant the narrower role to named roles instead. Test this first, because other Cortex features depend on `CORTEX_USER`.

## Whose permissions does an agent use?

Cortex Agents run with the caller's active role, set by the `X-Snowflake-Role` header or the user's default role, plus secondary roles. That role set must hold privileges on every object a tool touches: Search services, the tables behind semantic views, procedures and UDFs, and the warehouse.

Custom tools follow the rights model of the procedure. An owner's-rights procedure runs with the owner's privileges, which can be broader than the caller's. Review every owner's-rights procedure you expose as a tool.

By default, if the caller lacks a privilege, the agent skips that tool with a warning. Setting `orchestration.tool_not_accessible` to `reject` makes the run fail instead. For audited workflows, choose reject: a silent skip can produce a confident answer built from partial data.

## How do you control tools and MCP connectors?

Agents can use external MCP servers through MCP connectors. These are OAuth-only and per-user, and creating one needs the CREATE EXTERNAL MCP SERVER privilege. Keep that privilege with a platform team, the same way you control external access integrations.

The Snowflake-managed MCP server, which exposes Snowflake tools to outside agents, is limited to 50 tools and a recursion depth of 10.

For the wider MCP picture, see [how to govern MCP servers](/blog/how-to-govern-mcp-servers).

## How do you audit Cortex Agents?

Snowflake gives you three sources:

1. **Traces.** Agent traces go to `SNOWFLAKE.LOCAL.AI_OBSERVABILITY_EVENTS`, read with `GET_AI_OBSERVABILITY_EVENTS`. Reading needs MONITOR on the agent. Full, unredacted content needs the READ UNREDACTED AI OBSERVABILITY EVENTS TABLE privilege. Event rows cannot be modified.
2. **Usage.** `ACCOUNT_USAGE.CORTEX_AGENT_USAGE_HISTORY` records tokens, credits, role and inference region, with up to one hour of lag.
3. **Access history and query history** for the SQL that tools ran.

Snowflake's docs we checked do not state a retention period for observability events. Export them if your policy needs more than you can confirm. For cost reporting, see [how to track Snowflake Cortex AI spend](/blog/track-snowflake-cortex-ai-spend). For the grants a monitoring tool needs, see [Snowflake privileges for AI usage monitoring](/blog/snowflake-grants-for-ai-usage-monitoring).

## Where does inference run?

The account parameter `CORTEX_ENABLED_CROSS_REGION` controls whether Cortex may process requests in other regions. Only ACCOUNTADMIN can set it. Organizations created after March 9, 2026 default to `ANY_REGION`. If you have data residency commitments, check this setting on every account. The usage view's `inference_region` column shows where requests actually ran.

Cortex AI Guardrails, generally available since May 14, 2026, cover agents and CoWork, including indirect prompt injection in tool calls. They are turned on through the account-level `AI_SETTINGS` parameter.

## A Cortex Agents hardening checklist

1. Revoke broad Cortex grants from PUBLIC where you can, and grant `CORTEX_AGENT_USER` to named roles.
2. Keep MODIFY and OWNERSHIP on agents with a small platform role.
3. Review every owner's-rights procedure used as a tool.
4. Set `orchestration.tool_not_accessible` to `reject` for regulated workflows.
5. Restrict CREATE EXTERNAL MCP SERVER.
6. Set `CORTEX_ENABLED_CROSS_REGION` to match your residency policy.
7. Turn on Cortex AI Guardrails.
8. Export traces and usage history on a schedule.

Snowflake agents are one part of the estate. Compare [governing Databricks agents](/blog/govern-databricks-agents), and see the [cross-platform guide](/blog/how-to-govern-ai-agents-across-platforms) and [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane). To talk through your estate, [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**What privileges does a user need to run a Cortex Agent?**
USAGE on the agent, a Cortex database role, and privileges on each tool object through their active or secondary roles.

**What happens if the caller cannot access one of the agent's tools?**
By default the agent skips the tool with a warning. Set `orchestration.tool_not_accessible` to `reject` to fail the run instead.

**Where are Cortex Agent traces stored?**
In `SNOWFLAKE.LOCAL.AI_OBSERVABILITY_EVENTS`, read with `GET_AI_OBSERVABILITY_EVENTS`. Reading needs MONITOR on the agent.

## Key facts

Snowflake Cortex Agents (GA Nov 4, 2025) are schema-level AGENT objects with CREATE AGENT, USAGE, MODIFY, MONITOR and OWNERSHIP privileges. SNOWFLAKE.CORTEX_USER is granted to PUBLIC by default; SNOWFLAKE.CORTEX_AGENT_USER gives agent-only access. Agents run with the caller's active and secondary roles; inaccessible tools are skipped unless orchestration.tool_not_accessible is reject. Traces go to SNOWFLAKE.LOCAL.AI_OBSERVABILITY_EVENTS (MONITOR needed). CORTEX_AGENT_USAGE_HISTORY lags up to 1 hour. CORTEX_ENABLED_CROSS_REGION is ACCOUNTADMIN-only; organizations created after Mar 9, 2026 default to ANY_REGION. Snowflake Intelligence is now Snowflake CoWork. Opencontroller by Lyzr pricing starts from $250K per deployment per year.

## Related

- [What Snowflake privileges does an AI governance tool need to monitor Cortex usage?](https://opencontroller.lyzr.com/blog/snowflake-grants-for-ai-usage-monitoring.md)
- [How do you track Snowflake Cortex AI spend by agent or team?](https://opencontroller.lyzr.com/blog/track-snowflake-cortex-ai-spend.md)
- [How do you govern AI agents built on Databricks?](https://opencontroller.lyzr.com/blog/govern-databricks-agents.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
