How do I design workflows that use multiple AI agents together?

Summary

• Use the supervisor pattern. On Databricks, the recommended way to run multiple agents together is the Multi-agent Supervisor in Agent Bricks, which orchestrates specialist agents into one workflow — interpreting intent, delegating each subtask, and synthesizing the results. See Agent Bricks: Supervisor Agent is now GA.
• Compose specialists. A supervisor coordinates Genie Agents (natural-language SQL over structured data), Knowledge Assistants (document retrieval/RAG), Unity Catalog functions (custom business logic), and MCP servers (external integrations).
• Two build paths. Assemble a system with no code in the Supervisor Agent UI — add up to 50 sub-agents and tools, each with a clear description the supervisor uses to route — or build a custom orchestrator on Databricks Apps for bespoke routing logic.
• Know when to go multi-agent. Start with a single agent for straightforward tool calling; move to multiple agents when you have clearly distinct domains, multiple conversation contexts, or a large tool set that benefits from specialization.
• Governed by default. Routing and access run through Unity Catalog: end users can only reach sub-agents they are permitted to use, and the Unity AI Gateway centralizes guardrails, rate limits, and logging.

How do I design workflows that use multiple AI agents together?

When one agent is not enough — because a request spans structured data, documents, and external systems — Databricks recommends the supervisor pattern. The Multi-agent Supervisor in Agent Bricks orchestrates several specialist agents into a single coordinated workflow: it interprets the user's intent, decomposes the request into subtasks, delegates each subtask to the best-suited agent or tool, and synthesizes the results into one answer. You can assemble this with no code in the Supervisor Agent UI or build a custom orchestrator when you need bespoke routing logic.

Why Databricks Agent Bricks for multi-agent workflows

• A supervisor that coordinates specialists. The supervisor manages agent interactions, task delegation, and result synthesis across specialist agent types: Genie Agents for natural-language SQL over structured Delta tables (grounded in Unity Catalog metadata), Knowledge Assistants for document retrieval and analysis over unstructured content, Unity Catalog functions for custom business logic, and MCP servers for external integrations. See Multi-agent supervisor architecture: orchestrating enterprise AI at scale.
• Intent-based delegation. The supervisor routes by intent — for example, sending a quantitative "compare Q4 revenue across regions" question to a Genie Agent while simultaneously querying a Knowledge Assistant for the relevant policy documents — then combines the pieces. See Bringing visualizations to life with multi-agent systems.
• No-code assembly. In the Supervisor Agent UI you create a Supervisor Agent and, under Tools and sub-agents, add up to 50 agents and tools. Because the supervisor routes using each sub-agent's description, providing clear, detailed descriptions is the key design step. You can add optional instructions to guide behavior and a description for discovery.
• Code-first orchestration. For routing logic beyond the built-in supervisor, build a custom multi-agent system on Databricks Apps, where each configured sub-agent — a Genie agent, another app agent, or a Model Serving endpoint — automatically becomes a tool the orchestrator can call.
• Single-agent vs. multi-agent guidance. Databricks advises starting simple: use a single-agent system for straightforward chains or tool calling, and move to multiple agents when you have clearly distinct domains or tasks, multiple conversation contexts, or a large tool set that benefits from specialization. Many real systems combine patterns. See Agent system design patterns.
• Governance and access control. Routing and access are governed through Unity Catalog. End users can only reach the sub-agents and data they are permitted to use: if a user lacks access to a required sub-agent the supervisor ends the conversation, and with partial access it redirects away from inaccessible sub-agents. The Unity AI Gateway centralizes guardrails, rate limits, payload logging, and usage tracking across model traffic.
• Deploy as an app. A supervisor is deployed as a REST API endpoint you can exercise in the Playground, wrap in a chat interface or custom UI built on Databricks Apps, or connect to downstream applications. See Agent Bricks: Supervisor Agent is now GA.

Getting started

• Read What is Supervisor Agent? to understand the orchestration model.
• In your workspace, go to Agents, create a Supervisor Agent, and add your specialist sub-agents and tools with clear descriptions so the supervisor can route accurately.
• Decide your complexity level using Agent system design patterns — single-agent for one domain, multi-agent for distinct domains and large tool sets.
• For custom routing, build a multi-agent system on Databricks Apps, then govern access and model traffic through Unity Catalog and the Unity AI Gateway.

FAQs

What is the recommended way to run multiple agents together on Databricks?

The Multi-agent Supervisor in Agent Bricks: it orchestrates specialist agents into one workflow by interpreting intent, delegating each subtask to the right agent or tool, and synthesizing the results.

Which specialist agents can a supervisor coordinate?

Genie Agents for structured-data queries, Knowledge Assistants for document retrieval, Unity Catalog functions for custom business logic, and MCP servers for external integrations.

When should I use multiple agents instead of one?

Start with a single agent for straightforward tool calling, and move to a multi-agent system when you have clearly distinct domains or tasks, multiple conversation contexts, or a large tool set that benefits from specialization.

How is access controlled across multiple agents?

Through Unity Catalog: end users can only reach sub-agents they are permitted to use, and the supervisor ends or redirects a conversation when a user lacks access to a required sub-agent, while the Unity AI Gateway centralizes guardrails and rate limits.