What patterns work best for orchestrating multi-agent workflows?

Summary

• Start simple, then add complexity. Begin with deterministic chains for well-defined tasks, move to a single agent with tool calling, and adopt multi-agent systems only for clearly distinct domains or large tool sets.
• Use a supervisor (hierarchical) pattern for most enterprise workflows. A supervisor agent plans the task and delegates to specialized worker agents, which works well for stable, well-understood workflows and is easy to audit.
• Agent Bricks Supervisor Agent is a generally available, managed orchestration layer that coordinates Genie Agents, agent endpoints, Unity Catalog functions, MCP servers, and custom agents from one endpoint.
• Treat each subagent as a tool. The orchestrator interprets intent, routes to the right specialist in sequence or in parallel, and synthesizes results; production systems add retry logic, task queuing, and full observability.
• Govern and observe. Unity Catalog provides on-behalf-of permissions and audit trails, and MLflow traces every step, retrieval, and tool call.

What patterns work best for orchestrating multi-agent workflows?

The best pattern depends on how distinct your tasks are and how much control and auditability you need. A good rule is to start simple and add complexity only when it pays off: begin with deterministic chains for well-defined tasks, move to a single agent with tool calling, and adopt a multi-agent system only for clearly distinct domains or large tool sets. When you do go multi-agent, a supervisor (hierarchical) pattern is the most common choice for enterprise workflows. On Databricks, the Agent Bricks Supervisor Agent provides a managed way to implement it.

Why Databricks Agent Bricks for multi-agent orchestration

• A managed supervisor. Supervisor Agent is a generally available orchestration layer that coordinates Genie Agents, agent endpoints, Unity Catalog functions, MCP servers, and custom agents from a single endpoint, managing agent interactions, task delegation, and result synthesis.
• Two ways to build. Use a no-code UI that assembles components and deploys to a REST API endpoint in minutes, or an advanced option using LangGraph in Databricks Notebooks with agentic memory and Skills, integrated with Lakebase (Databricks serverless Postgres).
• Recognized orchestration patterns. Orchestration patterns trade off control, resilience, and scalability: hierarchical orchestration, where a supervisor plans and delegates to specialized worker agents (good for stable, well-understood workflows and auditability); decentralized orchestration, where agents self-organize peer-to-peer (resilient but harder to audit); and hybrid patterns that combine both in a single system. Teams with strict compliance needs typically favor centralized orchestration for auditability.
• Sequential and parallel routing. The orchestrator treats each subagent as a tool: it interprets the user's intent, queries specialized agents (for example, Genie for structured data and RAG agents for unstructured documents) in sequence or in parallel as needed, and then synthesizes the results. See building a multi-agent system.
• Tool-calling subagents. An orchestrator can connect to Databricks Apps agents (called through the Responses API), Genie Agents (via the built-in Databricks MCP server), and serving endpoints such as knowledge assistants, agents, or models on Model Serving. Each subagent is configured with a description that directly affects routing quality.
• Production-grade coordination. Robust multi-agent systems add automatic retry logic for transient failures, task queuing for high-volume workflows, and comprehensive observability with traces, logs, and metrics that show exactly what each agent did and why. See the agentic systems guide.
• A pattern-agnostic framework. Mosaic AI Agent Framework is generally available and is agnostic of pattern choice, making it easy to evolve your design as applications grow.
• Observability and evaluation. MLflow provides tracing across every step, retrieval call, and tool invocation, LLM-judge and synthetic-data evaluation, and ongoing monitoring with configurable trace sample rates.
• Enterprise governance. Unity Catalog provides on-behalf-of (OBO) permissions so agents act with the requester's scope rather than a shared service principal, an audit trail for every model and tool call, and access control for MCP servers and subagents. AI Gateway (Beta) adds governed tool access to external services. The supervisor has built-in access controls so users only reach subagents and data they are permitted to use.

Getting started

• Read orchestration patterns for multi-agent coordination to choose between hierarchical, decentralized, and hybrid designs.
• Review agent system design patterns for the start-simple-then-scale progression.
• Try the Supervisor Agent to coordinate Genie Agents, custom agents, and tools from one endpoint.
• Follow build a multi-agent system on Databricks to wire up subagents and routing.

FAQs

What is the most common multi-agent orchestration pattern?

The hierarchical (supervisor) pattern, where a supervisor agent plans the overall task and delegates to specialized worker agents. It suits stable, well-understood workflows and is straightforward to audit.

When should I use multiple agents instead of one?

Start with deterministic chains, then a single agent with tool calling. Move to a multi-agent system only when you have clearly distinct domains or a large set of tools that a single agent cannot manage cleanly.

How does a supervisor route work between subagents?

The supervisor treats each subagent as a tool and uses its description to route requests to the right specialist, in sequence or in parallel, then synthesizes the results. Clear subagent descriptions directly improve routing quality.

How are multi-agent systems governed on Databricks?

Unity Catalog enforces on-behalf-of permissions and audit trails for every model and tool call, and the Supervisor Agent has built-in access controls so users only reach subagents and data they are allowed to use.