Snowflake ups AI agent context, governance game

Snowflake launched a series of tools and features designed to make it easier for enterprises to deploy AI agents with governance and context.

The headliner at Snowflake Expedition 2026 was the company's new Business Ontology within Horizon Context. Snowflake's Horizon Context sits within Horizon Catalog. Business Ontology gives AI agents shared, pre-approved business definitions for accurate answers.

Business Ontology will ride along with Cortex AI Gateway, which governs how agents access models, tools and take action with visibility. Snowflake also said that Cortex AI Gateway extends Snowflake's security and governance to all agents.

Snowflake also announced Coco and CoWork Bots, a new class of AI agents that work in the background. Automations in CoCo and CoWork handle recurring work autonomously and retain and share skills, memory and artifacts.

• Snowflake Q2 growth surges on AI and data demand
• Constellation ShortList™ Metadata Management & Data Governance
• Constellation ShortList™ Semantic & Context Management for Analytics and AI

The new CoCo and CoWork Bots were part of updates to Snowflake CoCo, a coding agent for data teams, and Snowflake CoWork, a personal work agent.

Snowflake also announced autonomous data engineering tools including blueprints for self-healing pipelines in CoCo Cloud Agents and CoCo Automations. In addition, developers can move Apache Spark workloads into Snowflake with fewer code changes.

Other announcements from Snowflake include:

• Agent Trace Intelligence, which allows enterprises to query agent execution against business data and fix inefficient loops and failing tools.
• Cortex Sandboxes, which give agents isolated environments to run custom code with enterprise data in Snowflake.
• Cortex Training, a tool for enterprises to build AI models on their own data.
• Profiles in CoCo and Profiles in CoWork were added for role and team-specific context for AI experiences.
• Skills and Plugin Sharing to codify CoCo workflows and share them with governance.
• Snowflake Cortex AI Gateway will get auto model tiers with built in model routing to match the model with the task and optimize performance and cost.

The big picture

Constellation Research analyst Mike Ni said Snowflake Business Ontology will benefit developers as well as data engineers.

"When someone asks which customers need attention, the agent has to connect the customer to the relevant accounts, products, and activity. A useful ontology makes those connections reusable. The payoff for developers is less time stitching context together and less rework when an agent answers the wrong question."

“A governed ontology can reduce conflicting meanings across reports, applications, and agents, provided those systems actually use it. Keeping it useful requires business owners to own and approve the meaning and data teams who maintain the mappings and test changes."

The broader themes aside from context and ontology from Snowflake's announcements revolved around the governance and the observability needed for continuous improvement. Regarding Snowflake's Agent Trace Intelligence, Ni said:

"With several agents, models, and tools involved, a bad answer can start well upstream of where it appears. Developers need to follow the execution across handoffs to find the cause. Otherwise, they risk changing the prompt when the actual problem is a failed tool or missing data. For instance, Cost per model call tells you very little about the cost of getting a job done. One request can trigger several agents, retries, and tools before producing a usable result. Operators need to see that whole path and measure cost per successfully completed task."

"Traces explain how the agent ran. Business data helps establish whether the work accomplished anything. Connecting the two enables teams to compare the steps taken with the order resolved, the case reopened, or the approval still outstanding. Decision traces, feedback loops, and outcome measurement are becoming fundamental to operating AI agents reliably to then invest to improving them based on what actually works."