How do I scale custom AI agents across teams and business units?

Summary

• Build once, govern everywhere. Agent Bricks and the Agent Framework let teams build, evaluate, and deploy production agents grounded in governed enterprise data in the lakehouse.
• A single system of record for AI. Unity Catalog registers models, agents, MCP servers, tools, and skills with unified access control, discovery, lineage, and audit across the organization.
• Permissions travel with the agent. On-behalf-of-user authorization means agents respect the data permissions of the person invoking them, so access stays correct across business units.
• Orchestrate, don't rebuild. The Supervisor Agent (generally available) coordinates multiple agents and tools — Genie Agents, Knowledge Assistants, and MCP servers — behind one entry point.
• Enforce policy at runtime. Model Serving and the AI Gateway add rate limiting, payload logging, and usage tracking to every deployed agent.

How do I scale custom AI agents across teams and business units?

Scaling agents beyond a single team is less about building more models and more about governing, orchestrating, and reusing agents consistently. The Databricks Data Intelligence Platform provides an integrated path — agent authoring with Agent Bricks and the Agent Framework, one governance layer in Unity Catalog, and managed deployment through Model Serving and the AI Gateway — so many teams can ship agents that are discoverable, permissioned, and observable across the organization.

Why Databricks Agent Bricks helps you scale agents

• Production agent authoring. Agent Bricks builds custom agents grounded in governed enterprise data in the lakehouse, and the Agent Framework provides the build, evaluate, iterate, and deploy lifecycle with built-in metrics, custom LLM judges, and MLflow tracing.
• A unified registry for AI assets. Unity Catalog extends data governance to cover models, agents, MCP servers, tools, and skills in a single system of record, with access control, discovery, lineage, and audit.
• Governed, reusable tools. Unity Catalog functions can be registered as agent tools and shared across teams and popular agent-authoring libraries, so one team's governed tool becomes another team's building block.
• On-behalf-of-user authorization. Agents respect the underlying data permissions of the human invoking them, and role-based access controls apply to models, tools, and connections — keeping access correct as agents spread across business units.
• Orchestration without rebuilding. The Supervisor Agent, generally available, uses a dynamic supervisor pattern to coordinate Genie Agents for structured data, Knowledge Assistants for unstructured data, and MCP servers for tools — letting each team own and iterate on its own agent while users get a single entry point.
• Runtime enforcement and monitoring. Model Serving deploys agents as scalable endpoints with permissioning, rate limiting, payload logging, and usage tracking, and the AI Gateway evaluates model, tool, and agent calls against policy before execution.

Getting started

• Read Introducing new governance capabilities to scale AI agents and governing AI agents at scale with Unity Catalog.
• Build and evaluate an agent with the Agent Framework, and register Unity Catalog functions as governed tools.
• Connect structured data through Genie and orchestrate multiple agents with the Supervisor Agent.
• Deploy through Model Serving and apply organization-wide policy with the AI Gateway.

FAQs

How does Databricks keep agent access correct across business units?

Unity Catalog applies role-based access control to models, tools, and connections, and on-behalf-of-user authorization means an agent respects the data permissions of the person invoking it.

Can teams reuse each other's agents and tools?

Yes. Unity Catalog acts as a single registry for models, agents, MCP servers, tools, and skills, so assets one team builds are discoverable, governed, and reusable by others.

How do I coordinate multiple agents for end users?

The generally available Supervisor Agent orchestrates multiple agents and tools — Genie Agents, Knowledge Assistants, and MCP servers — behind a single entry point.