Snowflake opened Expedition 2026, its new virtual event that replaces BUILD, on 3 November 2026. The message was simple: AI agents only work if the data behind them is fresh, governed and well defined. Most announcements fall into three groups: agents that do data engineering for you, faster and more real-time data, and controls that make agents trustworthy. This guide sorts them by what you can use today and what is still coming. For the background, see our Snowflake vs Databricks for AI agents comparison.
Autonomous data engineering with CoCo
CoCo, Snowflake's coding agent for data work (formerly Cortex Code), was the centre of the launch. Until now it mostly ran on an engineer's machine. It can now run inside Snowflake, on a schedule, and inside your normal review process.
CoCo at Expedition 2026
Cloud Agents
CoCo runs in a Snowflake-managed container, so long jobs continue after you log off.
Automations
Schedule recurring work, such as a daily pipeline health check, described in plain language, with audit trails.
Restricted Session Scope
Limit what an agent is allowed to do in a session.
GitHub integrationNow
Proposed code changes go through your team's normal pull-request review.
BotsUpcoming
Background agents that only interrupt you when a decision is needed.
Snowflake for SlackUpcoming
Ask questions, start workflows and review results from Slack.
Pipelines, streaming and real-time analytics
Build pipelines
Workspaces- Pipeline Builder (public preview): connect notebooks, Python and SQL steps visually, or ask CoCo to draft the pipeline
- Each pipeline is a project you can review and keep in Git
Fewer hand-written task graphs for routine pipelines.
Streaming
Snowpipe Streaming- Elastic Channels (GA): streaming capacity grows automatically with data volume
- Up to 20 GB per second of uncompressed data per table
- Snowflake Datastream for Kafka-based apps is coming soon
- Streaming Transformations in SQL is in private preview
Data lands in seconds instead of minutes.
Serving
Interactive Analytics- GA: sub-second answers for apps, dashboards and AI agents straight from Snowflake
- Designed for many short queries at once, such as an app or an agent asking questions all day
One copy of the data for analysis and for live apps.
Spark on Snowflake
For teams with existing Apache Spark code, Snowpark Connect lets Spark jobs run on Snowflake with fewer changes. Support for Apache Iceberg tables, in Snowflake or in external catalogs, is now GA, and CoCo Spark code migration (GA) reviews Spark code and rewrites it for Snowflake. DataFrame Writer V2, flexible partitioning, a monitoring view for Spark jobs and Write-Audit-Publish on Iceberg tables, which lets you test a pipeline on a branch of the data before production, are in public preview. One customer quoted by Snowflake said moving heavy Spark workloads cut its total cost by 60%, a vendor-supplied figure worth testing on your own jobs.
Governance: making agents trustworthy
Snowflake CEO Sridhar Ramaswamy framed the second half of the launch around trust: an agent that acts on your business must use the right definitions, respect the same permissions as people, and stay within budget.
- Business Ontology
- An extension of semantic views that stores approved business concepts and metrics once, so CoCo, CoWork and other agents give the same answer to the same question. Private preview, public preview "soon".
- Cortex AI Gateway
- One entry point for agents, now adding abuse detection, tracing, sandboxes, AI cost controls, dynamic model routing and tool access rules from the Natoma acquisition.
- AI security
- Existing masking and access policies apply to agent actions, with activity monitoring and a record of what each agent did.
- Cortex Training
- Post-train open-weight models on your own data without managing GPUs. Public preview; Snowflake claims up to 61% lower cost than other serverless training services, based on their published prices.
- CoWork and Profiles
- CoWork, the agent for business users, can now run recurring analysis, and Profiles add role-specific context and admin controls to both CoWork and CoCo.
If this sounds familiar, it is the same idea we described in The semantic layer is back: models are not the bottleneck, business context is. Databricks is moving the same way, with relationships between metric views now in beta.
Also announced
- Oracle partnership: Snowflake will be offered as a first-party service on Oracle Cloud Infrastructure.
- AI Data Package: more than 60 curated industry datasets with semantic views and a ready CoWork agent.
- Agentic discovery on Snowflake Marketplace: find and install data products by asking in plain language.
- Integrations: CoCo and CoWork now work with Slack, Excel and Google Sheets, and through a mobile app.
What to do next
- 1
Try Interactive Analytics on one dashboard
It is GA. Point a slow, frequently used dashboard at an interactive warehouse and compare speed and cost.
- 2
Pilot a CoCo Automation
Pick one recurring check, run it with a restricted scope, and send results to a channel your team already reads.
- 3
Plan for Business Ontology
Write down your ten most important metric definitions now. Whoever has them agreed first will get the most from ontology-based agents.
- 4
Spark teams: test the migration
Run CoCo Spark code migration on one non-critical job and compare results and cost before committing.
Check your understanding
1/3Which CoCo feature lets work continue after you close your laptop?
Cloud Agents run CoCo in a Snowflake-managed container, and are generally available.
2/3What is Business Ontology for?
It extends semantic views so agents stop giving conflicting answers to the same question.
3/3Which announcement serves sub-second answers to apps and agents?
Interactive Analytics is GA and serves real-time answers directly from Snowflake.
You got 0 out of 3.
More on Snowflake: this week's data platform news, Snowflake secondary indexes, BigQuery vs Snowflake in 2026 and the Snowflake October 2026 update.
Sources: Snowflake press release "Snowflake Delivers the Foundation for Autonomous Data Engineering" (3 November 2026); Snowflake's Expedition 26 blog posts; SiliconANGLE and Investing.com coverage of Expedition 2026 (3 November 2026). Customer results and cost claims are Snowflake's; availability varies by region and edition.
Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 3 November 2026. About me