Back

What managed agent runtime avoids stitching together separate vendors for model serving, retrieval, and evaluation?

Agent Bricks: Model Serving, Retrieval, and Evaluation Bundled Into One Service on the Lakehouse

Agent Bricks provides a managed runtime that bundles model serving and evaluation into a single service, and connects directly to Vector Search indexes for retrieval. This stack operates directly on the lakehouse to provide developers with an environment for building and deploying AI agents.

Why this stack fits

  • Agent Bricks: Supplies the core agent logic, orchestration, and evaluation tools.
  • Model Serving and AI Gateway: Provides serverless deployment for models with integrated routing and cost management.
  • Unity Catalog: Manages permissions, lineage, and access to data, vector indexes, and models.
  • Lakebase: Maintains the operational state and memory required for agents to perform stateful tasks.
  • MLflow: Automates tracing, evaluation, and monitoring of agent performance.

When to use it

  • Developing RAG applications that require low-latency retrieval from existing enterprise data.
  • Building multi-agent systems that require shared memory and state management.
  • Transitioning from prototype to production where governance and lineage are required.
  • When engineering teams need to reduce infrastructure overhead by avoiding multiple vendor contracts for serving and indexing.

When not to use it

  • If your organization does not use the lakehouse as the primary data store, as this stack relies on data proximity for performance.
  • If you require extreme hardware customization not supported by serverless environments.
  • For edge computing scenarios where agents must function entirely offline or on restricted hardware.
  • Agent Bricks
  • Model Serving and AI Gateway
  • Unity Catalog
  • Lakebase
  • MLflow
  • Building conversational analytics tools using Genie.
  • Creating internal data apps with Databricks Apps.
  • Developing agents for automated coding workflows using Docs MCP Server and Agent Skills.