Snowflake used its virtual Expedition 2026 conference on 3 November to unveil a bundle of features aimed at one problem: getting AI agents to behave sensibly inside a large company. The package has three headline parts. Business Ontology is a way of fixing what company terms mean. An expanded Cortex AI Gateway controls what agents may reach and spend. And autonomous data engineering lets agents build, monitor and troubleshoot data pipelines with less human involvement.
Business Ontology is the newest piece of Horizon Context, which sits inside Snowflake's Horizon Catalog. It builds on the semantic views Snowflake already offers, letting an organisation record its approved definition of a term such as net revenue and tie it to the tables and views behind it. The division of labour is spelled out: the ontology governs what a term officially means, a Semantic View does the calculation, and a component called Cortex Sense guesses at meanings nobody has defined. Where an approved definition and an inference collide, the approved definition wins.
Two of Snowflake's own agents draw on those approved concepts. CoCo is a coding agent for data teams, and CoWork is a personal work agent. CoCo also handles migration: definitions that already live in a third-party catalog or a spreadsheet can be brought across in a single guided session. They arrive as drafts for the owners to approve, and every change is logged. Horizon Context connectors, meanwhile, pull schemas, query logs and dashboards from tools including Tableau, Power BI, dbt, SQL Server and PostgreSQL. Business Ontology is in private preview, with a public preview promised soon, so most customers cannot yet try it. The idea behind it is that a definition should have a named owner, a trail of changes and one official meaning that every agent and every person sees, instead of living in whichever document a team happened to keep.
“An agent that acts on your business has to be trusted, and it has to know three important things: what the data means, which data to trust, and what it's allowed to do”
Sridhar Ramaswamy, Snowflake CEO, at Expedition 2026 · siliconangle.com ↗
The gateway is the second big piece. Cortex AI Gateway first appeared in July 2026 as a central place to govern and watch agents. This update adds abuse detection, observability and tracing, Cortex sandboxes, AI cost governance and dynamic model routing, plus tool integration that uses technology from Natoma, a company Snowflake bought. One policy set applies to agent connections across clouds, frameworks and AI providers, and that includes calls to MCP servers. Agents inherit the controls already set for human users, so data masking and access restrictions follow them. Security and IT teams can also cap costs, choose models for efficiency or accuracy, and limit which outside tools an agent can touch. The point, as Snowflake describes it, is that a company running many agents from many vendors gets a single place to set rules rather than configuring each connection separately.
Customers named in the announcement give a sense of the use cases. Glean, Kargo and Uniphore are cited as users or partners of the gateway, while the marketing agencies dentsu and Merkle are named as users of Business Ontology. Kargo's data chief says the gateway answers a question the company could not previously settle: which teams and which agents were behind its AI spending, now with hard limits that can be set. Dentsu's Joe Tobey argues that once agents query data, definitions need an accountable owner and cannot stay in static data dictionaries.
Snowflake's chief executive, Sridhar Ramaswamy, frames the shift as a change in where advantage lies: it moves away from specialist developers and towards proprietary business knowledge, as people in business roles become developers themselves. His product chief, Christian Kleinerman, says governance is what earns trust and speeds adoption, and that the companies that come out ahead will be those whose agents grasp the business's meaning and whose actions can be accounted for. SiliconANGLE adds a practical illustration of the risk: without a shared definition, two agents can return incompatible answers from the same accurate data.
The company backs the pitch with a survey it sponsored through MIT Technology Review Insights. It polled 400 senior data and technology executives in July and August 2026. Some 55% named missing business context as the main reason agent insights disappoint, and 45% pointed to fragmented data. Snowflake, which says it has more than 14,500 customers, presents the new tools as the answer to both problems.
Two other items rounded out the day. Snowflake widened its partnership with Oracle so that its platform becomes a first-party service on Oracle Cloud Infrastructure, and Cortex Training entered public preview, letting customers tailor open-weight models with their own data without managing GPUs or infrastructure.
About Snowflake
Snowflake is a cloud AI data platform company, founded on 23 July 2012 in San Mateo, California, by Benoit Dageville, Thierry Cruanes and Marcin Zukowski. It is headquartered in Menlo Park, California, with a principal executive office in Bozeman, Montana since it became a distributed company in May 2021.
The company went public in September 2020 in an IPO that raised $3.4 billion, among the largest software IPOs. Frank Slootman retired as chief executive on 28 February 2024 and was succeeded by Sridhar Ramaswamy, a cofounder of Neeva. This is not the first time Snowflake has pitched agents: at its Summit conference on 2 June 2026 it presented an "agentic enterprise", adding Cortex Sense and Agent Identity, with Horizon Context enhancements built on its purchase of Select Star.
SiliconANGLE places the announcements in a wider race among enterprise software vendors to control the layer that sits between agents and company data, naming SAP's Autonomous Enterprise and Joule agents and Salesforce's Enterprise AI Harness and AI Control Plane as rivals in the same effort.