Databricks Agent Bricks

Agent Bricks is aimed at companies that want agents grounded in their own data. Rather than writing everything by hand, a team describes the task, such as answering questions over internal documents, extracting information from forms, or coordinating several agents, and the platform helps build it. It generates synthetic test data, evaluates quality with judges, and tunes the agent by trying different models and settings to improve accuracy and cost.

Because it sits on the Databricks platform, agents use data governed through Unity Catalog, so access controls follow the company’s existing rules. Observability and evaluation tools built on MLflow track how agents behave, and developers can bring their own frameworks. Lakebase gives agents a transactional database for state and application data, so applications can store memory and results next to the analytics data.

Agent Bricks is sold as part of Databricks’ usage-based pricing, so costs depend on the compute and models used. It makes the most sense for organizations that already store data in Databricks. Alternatives include Snowflake Cortex, Amazon Bedrock AgentCore, Google Vertex AI Agent Builder, and Microsoft Copilot Studio, as well as open-source frameworks like LangGraph. Agents still require careful evaluation before they are trusted with real workflows.