Skip to main content

Databricks Agent Bricks vs Snowflake Cortex Agents

Summary

  • Databricks Agent Bricks is the Databricks agent developer platform: you write agents in code with any framework or harness, reach frontier and open models through Unity Gateway, and govern data, tools, models and agents in Unity Catalog.
  • Snowflake Cortex Agents is a fully managed agentic platform where each agent is a schema-level Snowflake object that bundles a model, tools, orchestration settings and instructions.
  • Both ground agents in governed data: Cortex Agents through Cortex Analyst, Cortex Search and stored procedures, Agent Bricks through Databricks-managed MCP servers for Genie Agents, Databricks AI Search, Unity Catalog functions and SQL.
  • For quality, Agent Bricks uses open-source MLflow Tracing and evaluation with traces stored in Unity Catalog. Cortex Agents records traces as OpenTelemetry-style events and offers its own evaluation metrics.
  • We believe Agent Bricks is the right choice for custom, code-first agents over structured and unstructured data, and Cortex Agents suits teams whose data and SQL logic already live in Snowflake.

Databricks Agent Bricks vs Snowflake Cortex Agents

Both are managed platforms for agents that answer questions and take actions on governed enterprise data, and both let an agent pick tools and call them against that data. The difference is the build model. Cortex Agents is configured as an object inside Snowflake and works mainly through Snowflake tools. Agent Bricks is a developer platform: you bring agent code written with any framework or harness, and Unity Catalog governs everything the agent touches. We believe Agent Bricks is the right choice when you're building custom agents that combine tables, documents, external tools and your own logic.

What are Agent Bricks and Cortex Agents?

Databricks describes Agent Bricks as "the Databricks agent developer platform." Agents run on Agent Runtime, which hosts the agent server on Databricks Apps at a stable, authenticated endpoint. For teams that don't want to write code, Genie Agents offer a low-code way to build agents over tables and documents.
Snowflake describes Cortex Agents as "a fully managed agentic platform for building and running AI agents." An agent is "a schema-level object that bundles the agent's model, tools, orchestration settings, and instructions," shaped with natural-language planning and response instructions. You create one in Snowsight, with SQL commands, or through the REST API.

Agent Bricks vs Cortex Agents: how they compare

Capability Databricks Agent Bricks Snowflake Cortex Agents
How you build Agent code in any framework or harness; the Agent Bricks CLI (Beta) scaffolds and deploys projects An agent object created in Snowsight, with SQL commands, or through the REST API
Models Frontier and open models through one Unity Gateway API, switchable without rewriting the agent Models listed on Snowflake's model availability page; Snowflake recommends selecting auto
Structured data Genie Agents and the Databricks SQL MCP server Cortex Analyst, which generates SQL from natural language using a semantic view
Unstructured data Databricks AI Search Cortex Search
Custom tools Unity Catalog functions, plus external and custom MCP servers Stored procedures, user-defined functions and remote MCP servers, plus web search when enabled for the account
Conversation state Agent memory and sessions backed by Lakebase (Beta) Threads that persist conversation context across turns
Access control Unity Catalog permissions and lineage across data, tools, models and agents; tools set to user auth run with the requesting user's permissions The SNOWFLAKE.CORTEX_USER or SNOWFLAKE.CORTEX_AGENT_USER database role, plus privileges on the agent object and on the objects its tools use
Tracing MLflow Tracing, with traces stored in Unity Catalog as OpenTelemetry-format Delta tables OpenTelemetry-style event data covering planning, tool calls, responses and user feedback
Evaluation Built-in LLM judges, registered scorers and expert feedback Answer correctness, logical consistency, custom LLM-judge metrics, and tool selection and execution accuracy (Public Preview)
Where users reach it The invocation API, Responses API or SQL The REST API, Snowflake CoWork and Cortex Code

How Databricks approaches agents

Agent Bricks starts from the idea that agent quality depends on context, and context lives in the lakehouse. Through Databricks-managed MCP servers, a custom agent can hand a business question to a Genie Agent, run SQL, retrieve passages with Databricks AI Search, or call a Unity Catalog function, and each call is checked against Unity Catalog permissions.
Models stay swappable. Unity Gateway is a single control plane for model inference, tools and skills, with usage tracking, budgets, PII masking and audit logs. You can use Databricks-hosted frontier and open-weight models or route to your own capacity from other providers.
Quality runs on open source. MLflow Tracing records inputs, outputs, latency, token usage and cost for each step, and the same traces feed evaluation with LLM judges. Production monitoring, which is in Beta, runs those scorers on a sample of live traces.

FAQs

Can Agent Bricks agents use data that lives in Snowflake?

Yes, through Lakehouse Federation. Databricks can run federated queries on Snowflake data through a foreign catalog that mirrors the Snowflake database in Unity Catalog, so Unity Catalog governance applies to that data.

Is Agent Bricks generally available?

Databricks announced general availability of custom agents on Databricks Apps in April 2026. Some parts carry their own labels: the Agent Bricks CLI, Databricks Sandbox and production monitoring are in Beta.

Do I have to use a specific agent framework on Databricks?

No. Agent Bricks doesn't require a specific framework, and the CLI includes starter templates for LangGraph and the OpenAI Agents SDK.

Which should I choose?

If your agents mostly answer questions over data already modeled in Snowflake, Cortex Agents keeps the agent next to that data. If you need custom agent code, multiple frameworks, documents alongside tables, and open-source tracing and evaluation, we believe Agent Bricks is the right platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.