What is the best platform for creating agents with subagents?

Summary

• Agent Bricks Supervisor Agent is the managed platform for multi-agent orchestration. It analyzes requests, routes them to the right subagent, and synthesizes responses. Supervisor Agent is generally available.
• Subagents are specialized tools a supervisor delegates work to. These include Genie Agents, Model Serving endpoints, Unity Catalog functions, AI Search indexes, custom agents, and other Supervisor Agents.
• Routing accuracy and governance are critical. Supervisor Agent matches requests to subagents while Unity Catalog ensures users only access what they're permitted to use. Identity propagates through the full chain so every action is auditable.
• The platform supports up to 50 subagents per supervisor, covering data querying, external tools, and custom business logic. See Agent Bricks: Supervisor Agent is now GA.

What is the best platform for creating agents with subagents?

The best platform for building agents with subagents combines a managed orchestrator, a framework for tool connectivity, and governance that carries user identity through the multi-agent chain. On Databricks, this is delivered through Agent Bricks Supervisor Agent (the orchestrator), the Model Context Protocol (MCP) (the connectivity standard), and Unity Catalog (governance).

What are subagents?

Subagents are specialized agents or tools a supervisor delegates work to. One might query data through Genie, another might process that data, and a third might take action. A supervisor analyzes requests, determines which subagents can help, routes the work, and synthesizes results. This lets you break complex tasks into manageable pieces, where each is handled by the best tool.

Key capabilities to evaluate in a subagent platform

Capability	Why It Matters
Orchestration accuracy	The supervisor must understand user intent and route requests correctly. Poor routing wastes resources.
Subagent diversity	Can the platform connect to data agents, ML services, external tools, and custom logic?
Identity propagation	User permissions must travel through the multi-agent chain so users can't access unauthorized data.
Auditability	Every agent call, tool invocation, and data access must be logged for compliance.
Scalability	How many subagents can one supervisor coordinate? Can supervisors call other supervisors?

How Databricks approaches multi-agent orchestration

Supervisor Agent is a managed orchestrator for enterprise multi-agent systems. It takes requests, interprets intent, selects the right subagent (or combination), passes work and results between them, and returns unified answers. Supervisor Agent is generally available. See Agent Bricks: Supervisor Agent is now GA.
You can orchestrate up to 50 subagents, including Genie Agents, Model Serving endpoints, functions, and MCP servers. A supervisor can call Genie to query data, Model Serving to run inference, Unity Catalog functions as tools, AI Search for semantic search, published dashboards, other Supervisor Agents, and MCP servers (managed, external, or custom). Work flows through the chain: a supervisor might fetch data from Genie, route it through an external MCP server for enrichment, and synthesize results for the user.
Identity and permissions are enforced at every step. User identity propagates through the entire multi-agent chain. Each subagent respects user permissions, so users can only access authorized data and subagents. If a subagent is not accessible, the supervisor excludes it from routing. Every interaction is auditable.
Define subagents by natural language for intelligent routing. You describe each subagent: what it does, what inputs/outputs it has, when to use it. The supervisor uses this to make routing decisions. Provide example questions and feedback to refine behavior over time.
The Model Context Protocol standardizes agent-to-agent connectivity. MCP is an open standard letting agents connect to tools consistently. Databricks supports managed MCP servers (Databricks-hosted, governed access to Genie, AI Search, SQL, functions), external MCP servers (third-party tools via Unity Catalog with OAuth), and custom MCP servers (developer-built, hosted as Databricks Apps). See MCP on Databricks.

Use cases

• Customer support: Route questions to specialized agents for different domains, each with appropriate permissions.
• Data analytics: Query the lakehouse through Genie, route through external services, present enriched insights.
• Complex workflows: Orchestrate step-by-step automation (reports, campaigns, transactions) with agents handling each role.

FAQs

What is the difference between a supervisor and a subagent?

A supervisor is the orchestrator that analyzes requests, routes them, and synthesizes responses. A subagent is a specialized agent or tool that a supervisor delegates work to.

How does Supervisor Agent decide which subagent to use?

It matches the user's request to the subagent descriptions you provide. Describe each subagent (what it does, inputs, outputs). The supervisor uses this to make routing decisions.

Can subagents call other subagents?

Yes. Supervisor Agents can call other Supervisor Agents, enabling nested hierarchies to break large problems into smaller ones.

How are permissions enforced across multiple agents?

User identity propagates through the full multi-agent chain. Each subagent respects user permissions. Unity Catalog centralizes access control and audit.