Skip to main content

What is Agent Bricks?

Agent Bricks is the Databricks developer platform for building and deploying custom agents that power your agentic products and workflows. You write the agent in code, with any framework or harness and any model. Agent Bricks provides the managed infrastructure the agent needs in production: hosting, durable execution, memory, tools, model access, tracing, and governance.

To build and deploy your first agent, see the Agent Bricks quickstart. For full product documentation, see the Databricks agents docs.

What the platform gives you

Each building block is usable on its own, and the Agent Bricks CLI wires them together for a new project. The following diagram shows how a deployed agent uses them:

Agent architecture: client requests, runtime, Unity Gateway, memory, Sandbox, tracing, and evaluation feedback

You want toUse
Deploy your agentThe agent runtime hosts any framework or harness, stateful or stateless. DurableAgentServer serves it with synchronous, streaming, and background runs, and recovers runs that a crash or restart interrupts.
Give your agent contextManaged memory and sessions for conversation history and long-term memory, plus Databricks-managed and external MCP servers for tools and data.
Run code safelyDatabricks Sandbox gives the agent an isolated environment for the code it writes, with scoped access to governed data.
Connect to modelsUnity Gateway gives one API for frontier and open models, so you can switch models without changing agent code.
Debug, evaluate, and monitor your agentMLflow Tracing records each step the agent takes, locally and in production. MLflow evaluation scores the agent's quality on test data before you deploy and on production traces after.
Govern what your agent can doUnity Gateway governs access to models, MCP servers, and skills, with guardrails, rate limits, and usage tracking. Unity Catalog governs the data the agent reads.

Agent lifecycle

Building a production agent is iterative. Start with the simplest agent that works, measure it, and improve it in small steps:

  1. Start a project. Scaffold a new agent from a framework template with agentbricks init, or bring an agent you already built with LangGraph or the OpenAI Agents SDK. See Agent Bricks CLI.
  2. Write the agent loop. Put your prompts, tools, and orchestration in the framework code. Call models through Unity Gateway, so you can switch models later without changing the agent.
  3. Give the agent context. Add memory and sessions for conversation history and long-term memory, and add tools such as MCP servers, Genie Agents, and Unity Catalog functions with agentbricks tools add.
  4. Run and debug locally. Run the agent with agentbricks dev, send it real requests, and read its traces to see each model and tool call.
  5. Evaluate. Collect example requests and expected answers, then evaluate the agent with MLflow to measure the effect of each change to prompts, tools, or models.
  6. Deploy. Deploy the agent to the agent runtime with agentbricks deploy, which creates the stores the agent declares and grants the agent access to them.
  7. Monitor and improve. Review production traces, monitor quality with scorers, and add what you learn to your evaluation data. Then repeat from step 2 and redeploy.

Agent Bricks, AppKit agents, and Omnigent

Databricks offers three ways to work with agents. They solve different problems, and they combine:

Agent BricksAppKit agents pluginOmnigent
Use it whenYou're building a production custom agent. The agent is a product of its own, or many apps and services share it.You need lightweight model capabilities inside one app. The agent is a feature of the app, such as an assistant over the app's data.You want coding agents to help you build. Use them to write an app or an agent.
What it isA platform for building and deploying standalone custom agentsLightweight agents that you declare in your AppKit app's codeOne interface for Codex, Claude Code, Cursor, and other coding agents
Where it runsIts own Databricks app on the agent runtimeIn your app's runtime, at the app's built-in routesOn your machine or in the cloud
Who uses itAny app, service, user, or agent, through its HTTP APIYour app's users, through the app's UIYou and your teammates, while you build
  • For a lightweight agent inside an app you're building, see Agentic features. When other apps and services need to call the agent, or it needs its own deployment, build it with Agent Bricks instead. An AppKit app can call an Agent Bricks agent through its HTTP API.
  • To build an Agent Bricks agent with coding agents, use Omnigent as your development environment. Omnigent doesn't host the agent: you deploy it to the agent runtime.

For how all the Databricks developer services fit together, see Platform overview.

Where to next

Databricks Developer Hub

Ready to ship your next agentic app in minutes?

Read docs