For years the Snowflake versus Databricks debate was about warehouses versus lakehouses, SQL versus Spark. In 2026 it is about something else: which platform your AI agents run on. At their June conferences both companies announced agent products, semantic "context" layers and new governance for machine identities. This guide lays out what each offers today, in plain terms, so you can decide what matters for your team.
The two strategies in one picture
Snowflake
Agents included- Business users
- CoWork
- Builders
- CoCo (formerly Cortex Code)
- Context
- Cortex Sense, semantic views
- Models
- Claude first, plus others
- Ready-to-use agents in Slack, Excel and mobile
- Strong governance: each agent gets its own identity and permissions
- Agents and context tied to Snowflake
Best when you want agents working for business users quickly.
Databricks
Build your own- Business users
- Genie, Genie One
- Builders
- Agent Bricks, Genie Code
- Context
- metric views, Genie ontology
- Models
- wide roster via Unity AI Gateway
- Many models from many vendors, with spend caps and tracing
- Lakebase, a transactional database built for agents
- More assembly required
Best when you have engineers building custom agents.
Snowflake: what was announced
- CoWork
- The renamed Snowflake Intelligence: a personal agent for knowledge workers with research, memory, reusable skills and connectors to Google Drive, Salesforce and Slack. Available in Slack, on iOS and inside Excel.
- CoCo
- The renamed Cortex Code: a coding agent for data work that can run tasks on its own, on a schedule, or from VS Code, desktop, mobile and Slack.
- Cortex Sense
- A context layer that combines query history, table metadata, BI definitions from Power BI and Tableau, and semantic views, and serves the right business context to agents.
- Semantic views and Autopilot
- Governed business definitions inside Snowflake, with Autopilot creating and refining them automatically (preview).
- AI Agent Identity
- Each agent gets its own cryptographic identity, role-based permissions and data masking, instead of borrowing a user's access.
- Natoma
- An MCP gateway Snowflake is acquiring, to control which enterprise tools agents can call.
Without business contextWith Cortex Sense
Snowflake's reported figures from Summit 2026. Snowflake's own internal tests on complex queries showed 47% → 83%. Vendor results; test on your own questions.
Databricks: what was announced
- Agent Bricks
- A framework for building, evaluating and improving agents on your own data, now focused on code-first development.
- Genie and Genie One
- Natural-language analytics for business users, with an ontology layer giving agents business context.
- Unity AI Gateway
- One place to govern AI use inside and outside Databricks: model routing, MCP management, tracing, identity, and enforced spend caps by user, team or use case.
- Document intelligence
- SQL functions such as
ai_parse_document,ai_extractandai_classifyturn PDFs and images into tables. - Lakebase
- A transactional, Postgres-style database designed for agents and apps, next to the lakehouse.
- Metric views
- Databricks' semantic layer, now with rolling and cumulative measures.
Head to head
| Snowflake | Databricks | |
|---|---|---|
| Agent for business users | CoWork | Genie, Genie One |
| Agent for data builders | CoCo | Genie Code, Agent Bricks |
| Business context | Cortex Sense, semantic views | Metric views, Genie ontology |
| Model choice | Claude by default, others available | Many vendors through Unity AI Gateway |
| Agent identity | Per-agent identity and permissions | Identity integrations in the gateway |
| AI spend control | Monitoring | Enforced spend caps |
| Tool access (MCP) | Natoma gateway (acquisition) | MCP management in Unity AI Gateway |
| Open table format | Apache Iceberg v3 | Delta Lake and Iceberg |
How to choose
- 1
Start from who will use the agents
If business users need answers in Slack and Excel next quarter, Snowflake's packaged agents get there faster. If engineers are building custom agents, Databricks gives more building blocks.
- 2
Check where your data already lives
Moving data to follow an agent product rarely pays off. Both platforms read Iceberg tables, which keeps options open.
- 3
Keep definitions portable
Store metrics and business logic in open formats, such as dbt models and the Open Semantic Interchange standard, so they can feed whichever agent you use.
- 4
Test accuracy on your questions
Take 30 real business questions with known answers and run them through each platform's agent before deciding.
- 5
Budget for governance
Per-agent permissions, audit logs and spend limits matter more than the demo.
Check your understanding
1/3Which Snowflake product is aimed at business users rather than data builders?
CoWork (formerly Snowflake Intelligence) is the knowledge-worker agent; CoCo (formerly Cortex Code) is for building.
2/3What is a key strength of Databricks' Unity AI Gateway?
It routes and governs AI use across many model providers, with spend caps by user, team or use case.
3/3What is the best protection against lock-in for agent context?
Definitions in dbt and open standards such as OSI can feed any platform's agents.
You got 0 out of 3.
Weekly updates: data platform news, 21 to 27 October. Read more on why context makes or breaks AI answers in The semantic layer is back: context for AI agents, and the latest Databricks changes in Databricks: October 2026 update.
Sources: Atlan's Snowflake Summit 2026 announcements summary and Cortex Sense explainer; PointFive's review of Snowflake and Databricks Summits 2026; Databricks October 2026 release notes. Product availability (preview or general availability) varies by feature and region; accuracy figures are vendor claims.
Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 28 October 2026. About me