Governed Orchestration Architecture

A comprehensive guide to building governed AI applications on Databricks. Understand how authentication flows, MCP integration, Model Serving, and AI Gateway work together for secure, scalable orchestration.

The "mother of all" scenarios: stitching together Databricks services with proper auth flows

Agent Authentication MCP External Services Model Serving OAuth Federation Agent Memory

↓ Explore the architecture

Architecture Components

All the pieces that come together in a governed AI application. Each component has specific authentication requirements.

Entry Points

🌐

External App

Token Exchange

📱

Databricks App

Native Auth

↓

Model Serving (Deployment)

🤖

Custom Models

Agents, MLflow

🧠

Foundation Models

Hosted LLMs

🌉

AI Gateway

External LLMs

Authentication Methods (Agent → Resources)

⚡

Automatic

Passthrough

👤

OBO

User Identity

🔑

Manual

Credentials

↓

MCP Servers & Resources

🔮

Genie

Analytics

🔍

Vector Search

Retrieval

⚙️

UC Functions

Tools

📊

DBSQL

Queries

🔗

External MCP

UC HTTP Conn

📦

Custom MCP

Databricks Apps

🗃️

Lakebase (Postgres)

Agent Memory • OLTP State • OAuth/Native Auth

Governance Layer

🛡️

Unity Catalog

Permissions, ABAC, Row Filters, Column Masks, Lineage

Three Authentication Scenarios

Scenario A: External App → Databricks

Your application runs OUTSIDE Databricks and needs to call Databricks APIs.

OAuth Token Exchange — IdP token → Databricks token (requires Federation Policy)
Native OAuth with Entra — Azure AD direct integration
Server-side handling — Tokens managed on backend, never exposed to browser

Scenario B: Agent → Resources

Your agent runs ON Databricks (Model Serving) and needs to access resources.

Automatic Authentication Passthrough
On-Behalf-Of-User (OBO)
Manual Authentication
Declared at logging time

Scenario C: Databricks App → Resources

Your web app runs ON Databricks (Apps platform) and needs to access resources.

Native OAuth (automatic)
Unity Catalog integration
Databricks SQL access
Streamlit, Dash, Gradio, React

📚 Databricks Apps Docs →

🗃️ Lakebase: Managed Postgres for AI Applications

Lakebase is a fully-managed Postgres OLTP database. Register in Unity Catalog for unified governance, permissions, lineage, and cross-source analytics alongside your lakehouse data.

🧠 Agent Memory

Short-term (session) & long-term (cross-session) memory with checkpointing

⚡ Online Feature Store

Low-latency feature serving for real-time ML models

💾 OLTP State

Fast transactional data access for apps and agents

OAuth (U2M / M2M)

1-hour tokens as DB password
User, Group, or Service Principal
Requires SSL connection

Native Postgres Password

For apps that can't rotate tokens
Standard CREATE ROLE
Enable via instance settings

🛡️ UC Registration 🧠 Agent Memory ⚡ Online Features 🔐 Auth Docs

Governed Orchestration Architecture