Overview
What is Agent Bricks?
Agent Bricks is Databricks' enterprise agent platform for building, deploying, and governing agents that operate on your business data. It unifies model access, execution, governance, and context across a single system: from the model you call, to the data your agent reads, to the identity it acts under. In your workspace you configure Knowledge Assistants, Supervisor Agents, and custom Python agents. Databricks handles evaluation, tuning, and quality improvement, then hosts each agent at an HTTP endpoint your app can call.
Your AppKit app connects to Agent Bricks capabilities through two plugins: the Model Serving plugin for agents, foundation models, and governed endpoints, and the Genie plugin for natural-language queries over Unity Catalog tables.
How it fits together
Your AppKit app calls Agent Bricks through a Model Serving endpoint (a foundation model, Knowledge Assistant, Supervisor Agent, or custom Python agent) or a Genie Agent (natural-language queries over Unity Catalog tables). The Model Serving plugin and Genie plugin cover both.
AppKit plugins for Agent Bricks
| You want to | Use this plugin | Frontend helper |
|---|---|---|
| Call a foundation model (LLM) with chat messages | serving | useServingStream, useServingInvoke |
| Call an agent endpoint (Knowledge Assistant, Supervisor Agent, custom Python) | serving | useServingStream, useServingInvoke |
| Give users natural-language queries over Unity Catalog tables | genie | GenieChat, useGenieChat |
Pick the plugin that matches the resource. No other primitive is required for the AI surface.
Auth
Serving and Genie HTTP routes run on behalf of the authenticated user by default. If the user doesn't have CAN QUERY on the serving endpoint or CAN RUN on the Genie Agent, the call fails with a 403. You don't write the permission check.
For server logic outside a route handler, call AppKit.serving("alias").asUser(req).invoke(...) to keep the same behavior.
Why AppKit instead of raw fetch
You could call a serving endpoint directly with fetch and a token. The plugin isn't doing something you can't do yourself. It's doing these things so you don't have to:
- Routes run as the authenticated user, so per-user permissions apply automatically. Your users only see endpoints and data they're already allowed to see. No OAuth code on your side. See Execution context for the details.
- All streaming is handled for you. SSE parsing, abort on unmount, token accumulation, and error handling.
useServingStreamanduseGenieChatdo this. - No secrets in the frontend. The plugin proxies through your server and tokens stay on the backend. No PAT in the React bundle.
- When your serving endpoint publishes an OpenAPI schema, AppKit generates typed endpoint aliases with TypeScript types for request and response per alias. Autocomplete for chunk shapes, not
unknown.
Creating a custom agent is a Python workflow: the ResponsesAgent interface, an agent framework (OpenAI Agents SDK, LangGraph, LlamaIndex), and MLflow for tracing. See Author an AI agent on docs.databricks.com for that track.
Pick a template to start from
Start from a template that matches your use case. Each one includes the Model Serving or Genie plugin wiring, an app.yaml resource binding, and a working UI you can adapt.
| You want to... | Template |
|---|---|
| Add a streaming chatbot to your app | AI Chat App |
| Let users query tables in natural language | Genie Analytics App |
| Add multi-agent Genie switching to an existing app | Genie Multi-Agent Selector |
Where to next
- Unity AI Gateway for governed access to models, agent endpoints, and external tools.
- Genie Agents for chat-with-your-data over Unity Catalog tables.
- Custom agent endpoints for wiring Knowledge Assistant, Supervisor Agent, or your own Python agent.