Snowflake Adds Governance for AI Agents

At Expedition 2026 on Nov. 3, 2026, Snowflake announced product updates intended to govern business context, agent actions and AI spending across its platform. The company said the capabilities connect governed data, security controls, observability and cost management, with an open data foundation.

Snowflake Business Ontology, marked for public preview soon, adds a shared place to define business concepts independently of specific data structures. Cortex Sense, also planned for public preview soon, can find context across data and suggest definitions; designated human owners approve official definitions. Snowflake Semantic Views then apply approved logic to metrics. Organizations can import existing glossary definitions from third-party catalogs or spreadsheets in a single session using Snowflake CoCo, Snowflake’s coding agent. The guided import identifies domains, candidate terms, relationships and linked data assets, then marks them as drafts for owner approval. Approved definitions connect to semantic views, tables and views across teams; intentionally different definitions can remain approved with labels specifying where and when they apply.

Snowflake said Cortex AI Gateway, in public preview, provides a centralized, governed entry point for agents to access models and data. It applies consistent policies across supported clouds, frameworks and AI providers, and uses existing identity and access controls so teams do not need a second set of agent credentials. Cortex Sandboxes, in private preview, provide a private environment for agents to run custom code alongside Snowflake data while retaining the permissions of the person running the agent. They launch in seconds and shut down automatically when work is complete. Cortex AI Gateway Tools, in public preview, checks agent tool calls against policy using the end user’s identity and can block unapproved API calls, unauthorized actions and compliance violations. Cortex AI Guardrails, also in public preview, adds account-wide runtime protection against prompt injection, jailbreaks and novel attacks for agents routed through the Gateway, including third-party agents.

Snowflake’s generally available Observability feature records model calls, tool invocations, reasoning steps and guardrail decisions in governed tables in the customer’s account; traces can be queried with SQL and joined with business data. Agent Trace Intelligence, in private preview, uses these records to identify agents looping through costly work, repeatedly failing tools and the cost of each completed task. Generally available Cost Governance attributes spending to a team, department, agent or application and enforces hard limits in near real time. Dynamic Model Routing, in private preview, routes requests to organization-approved models based on cost, accuracy and latency. In internal testing, Snowflake said it delivered up to 3x greater token efficiency than a frontier-model-only path at the same quality.

Observe by Snowflake is a unified observability platform for investigating failures, agent quality and cost. Its Snowpipe Streaming ingestion pipeline, in private preview, writes logs, metrics and traces directly into Snowflake tables. Snowflake Data Share into Observe and the O11y Agent in Slack are also in private preview; the former creates zero-copy Observe datasets from customer data already in Snowflake, while the latter automates an initial response to issues and shares root-cause analysis in Slack. Agent Observability, in private preview, tracks performance, errors, token use and cost for agents and LLM applications, and supports searching and debugging traces, sessions and conversations. It runs online LLM-as-judge evaluations against production traffic, collects telemetry through OpenTelemetry and retains it in Apache Iceberg tables.

Snowflake said Business Ontology and Cortex AI Gateway are built on an open architecture using Apache Iceberg for storage, Apache Polaris for governance and Apache Ossie, which is incubating, for semantics. Horizon Catalog, built on Apache Polaris, offers a universal catalog for multicatalog and multiengine setups. Its generally available external-engine access lets compatible engines apply governance policies to Iceberg REST Catalog tables without migration. Audit history now logs interactions by Snowflake and external engines with Snowflake-managed and externally managed Iceberg tables when Horizon Catalog serves as the.

At Expedition 2026 on Nov. 3, 2026, Snowflake announced product updates intended to govern business context, agent actions and AI spending across its platform. The company said the capabilities connect governed data, security controls, observability and cost management, with an open data foundation.

Snowflake Business Ontology, marked for public preview soon, adds a shared place to define business concepts independently of specific data structures. Cortex Sense, also planned for public preview soon, can find context across data and suggest definitions; designated human owners approve official definitions. Snowflake Semantic Views then apply approved logic to metrics. Organizations can import existing glossary definitions from third-party catalogs or spreadsheets in a single session using Snowflake CoCo, Snowflake’s coding agent. The guided import identifies domains, candidate terms, relationships and linked data assets, then marks them as drafts for owner approval. Approved definitions connect to semantic views, tables and views across teams; intentionally different definitions can remain approved with labels specifying where and when they apply.

Snowflake said Cortex AI Gateway, in public preview, provides a centralized, governed entry point for agents to access models and data. It applies consistent policies across supported clouds, frameworks and AI providers, and uses existing identity and access controls so teams do not need a second set of agent credentials. Cortex Sandboxes, in private preview, provide a private environment for agents to run custom code alongside Snowflake data while retaining the permissions of the person running the agent. They launch in seconds and shut down automatically when work is complete. Cortex AI Gateway Tools, in public preview, checks agent tool calls against policy using the end user’s identity and can block unapproved API calls, unauthorized actions and compliance violations. Cortex AI Guardrails, also in public preview, adds account-wide runtime protection against prompt injection, jailbreaks and novel attacks for agents routed through the Gateway, including third-party agents.

Snowflake’s generally available Observability feature records model calls, tool invocations, reasoning steps and guardrail decisions in governed tables in the customer’s account; traces can be queried with SQL and joined with business data. Agent Trace Intelligence, in private preview, uses these records to identify agents looping through costly work, repeatedly failing tools and the cost of each completed task. Generally available Cost Governance attributes spending to a team, department, agent or application and enforces hard limits in near real time. Dynamic Model Routing, in private preview, routes requests to organization-approved models based on cost, accuracy and latency. In internal testing, Snowflake said it delivered up to 3x greater token efficiency than a frontier-model-only path at the same quality.

Observe by Snowflake is a unified observability platform for investigating failures, agent quality and cost. Its Snowpipe Streaming ingestion pipeline, in private preview, writes logs, metrics and traces directly into Snowflake tables. Snowflake Data Share into Observe and the O11y Agent in Slack are also in private preview; the former creates zero-copy Observe datasets from customer data already in Snowflake, while the latter automates an initial response to issues and shares root-cause analysis in Slack. Agent Observability, in private preview, tracks performance, errors, token use and cost for agents and LLM applications, and supports searching and debugging traces, sessions and conversations. It runs online LLM-as-judge evaluations against production traffic, collects telemetry through OpenTelemetry and retains it in Apache Iceberg tables.

Snowflake said Business Ontology and Cortex AI Gateway are built on an open architecture using Apache Iceberg for storage, Apache Polaris for governance and Apache Ossie, which is incubating, for semantics. Horizon Catalog, built on Apache Polaris, offers a universal catalog for multicatalog and multiengine setups. Its generally available external-engine access lets compatible engines apply governance policies to Iceberg REST Catalog tables without migration. Audit history now logs interactions by Snowflake and external engines with Snowflake-managed and externally managed Iceberg tables when Horizon Catalog serves as the.