Nov 17, 2026

Native Identity for AI Agents

Native Identity for AI Agents

TechStar Group · Snowflake Center of Excellence · Nov 16, 2026

AI agents now write and run SQL, build pipelines and kick off jobs, often on a schedule with nobody watching. That leaves every Snowflake team with two questions. How do you know when an agent, not a person, is doing the work? And how do you stop it from doing more than it should? Over 2026, Snowflake has answered both inside the platform.

The problem with borrowed credentials

Until recently, an agent either ran inside a person's session or used a shared service account. Either way, it inherited every privilege of whatever it logged in as. An assistant working for an account administrator was, in effect, an account administrator. Auditing had the same blind spot: a query an agent ran under someone's login looked exactly like one that person typed.

Knowing an agent is present

Snowflake now marks a session as agentic based on how it was opened. Sessions through Snowflake's own agents, such as Cortex Agents, Cortex Code and Snowflake CoWork, or through a Snowflake-managed MCP server, are flagged automatically. Third-party agents can be flagged too, by setting IS_AGENTIC to true on their OAuth integration, by using OAuth token exchange, or by signing in as a dedicated agent user. One catch: a programmatic access token on its own doesn't count, so a custom agent using one stays invisible until it moves to a supported path.

Delegated or autonomous

With agents visible, you decide how each one gets its access. A delegated agent, like a coding assistant helping an engineer, works as that person and draws on their role, ideally with a cap on top. An autonomous agent, like a scheduled job that watches pipelines overnight, gets its own identity.

That identity is the SERVICE_AGENT user type, generally available since July 2026. It can't have a password and every session it opens counts as agentic. The best way to sign it in is workload identity federation: the agent presents a short-lived token from the cloud it already runs on, so there's no secret to store, leak or rotate. The user type doesn't restrict anything by itself, though. The role you grant decides what the agent can reach, so it should be scoped tightly to the job.

Limits on what agents can do and see

Restricted Session Scope, generally available since September 2026, puts a privilege ceiling on agent sessions. It only ever removes privileges and never adds them. An administrator can apply it across the account so that agents can read data but never change it, even when the person behind them holds an admin role.

Masking and row access policies handle visibility. Because they can check whether an agent is active, a person can still see a sensitive column while an agent acting for them sees it masked.

An audit trail that writes itself

Query history now records which type of agent ran each query, and access history lists every agent involved in each data access. No extra logging pipeline is needed.

Why it matters

Most AI governance is added after the fact, through a proxy in front of the database, a separate logging service or conventions someone has to police. Snowflake builds it in instead. Agent identity, privilege ceilings, masking and audit sit on the same platform as the data and use the same roles and policies teams already manage. They also apply equally to Snowflake's own agents and to third-party ones. The result is less to build, fewer long-lived secrets and one place to answer the question that matters most: what did this agent actually do?

‍