Which platform is best for creating multi agent orchestration with metadata based agent descriptions?

Summary

• Metadata-based agent descriptions give an orchestrator written text to route work. Each agent or tool has a name and a plain-language description, and the orchestrator reads them to decide who handles a step. Databricks Supervisor Agent uses the description you write for each tool or sub-agent this way.
• No single platform is best for every team, so choose by who builds the agents. For new agents written in code, Databricks recommends the Agent Bricks CLI (Beta). For new agents built without code, it recommends Genie Agents.
• Agent Bricks hosts agents built with any framework or harness, and the CLI scaffolds LangGraph and OpenAI Agents SDK projects.
• Supervisor Agent documents description-driven orchestration of up to 50 agents and tools, but Databricks no longer recommends it for new agents.
• The supervisor limits each end user to the subagents and data that user can already access, and MLflow Tracing records every intermediate step.

Which platform is best for creating multi agent orchestration with metadata based agent descriptions?

Pick the platform where agent descriptions, permissions and traces sit next to the data and tools your agents use, and where your team can build the way it prefers. Databricks recommends the Agent Bricks CLI (Beta) for agents written in code, which scaffolds LangGraph or OpenAI Agents SDK projects, and Genie Agents for agents built without code. Open-source frameworks such as LangGraph and CrewAI also orchestrate multiple agents. The right choice depends on team skills, governance needs and how much orchestration code you want to own.

What are metadata-based agent descriptions?

In a multi-agent system, an orchestrator decides which agent or tool handles each step. Metadata-based descriptions give it text to decide with. An MCP tool definition has a name, a human-readable description and a JSON Schema for its inputs. A CrewAI agent is defined by a role, a goal and a backstory, and it can be allowed to delegate tasks to other agents. In Databricks Supervisor Agent, you write a description for each tool or sub-agent, and the supervisor uses it to coordinate them. The docs advise giving as much detail as possible, because that improves task delegation.

Key capabilities to evaluate

Capability	Why it matters
Description-driven routing	The orchestrator delegates well only when each agent and tool has a clear written description.
Mixed agent and tool types	Real workflows combine analytics agents, retrieval, functions and external tools.
Permission-aware delegation	Users should only reach agents and data they are already allowed to use.
Framework choice	Teams should keep the framework or harness they already know.
Tracing	A run that crosses several agents needs a record of each step to explain a wrong answer.
Governance of models and tools	Central control over which models and MCP servers agents can call.

How does Databricks approach multi-agent orchestration?

Agent Bricks is the Databricks agent developer platform, and Agent Runtime hosts agents built with any framework or harness. With the Agent Bricks CLI (Beta), agentbricks init --framework langgraph my-agent scaffolds a LangGraph project, agentbricks dev runs it locally and agentbricks deploy my-agent deploys it. agentbricks tools add genie-agent gives your agent Genie Agents as tools, and agentbricks tools add mcp adds a built-in MCP service and records it in the project's agent.toml file.
Genie Agents is the low-code route. Omnigent (Beta on Databricks) is a meta-harness for composing and collaborating with agents, such as combining coding agents, and you can use it to develop custom agents that run on Agent Bricks.
Supervisor Agent is the description-driven option. It coordinates Genie Agents, agent endpoints, Unity Catalog functions, MCP servers and custom agents. You select up to 50 agents and tools, describe each one, and can add instructions that tell the supervisor how to respond. Databricks lists it under "Other ways to build agents" and says earlier products such as Supervisor Agent are no longer recommended for new agents, so check the current docs first. Managing supervisors with the Databricks SDK for Python is in Beta.
The supervisor limits end users to the subagents and data they already have access to. If a user can reach no subagent, it ends the conversation, and if the user can reach some, it redirects the conversation away from the rest. Each type needs its own grant, such as EXECUTE on a Unity Catalog function or CAN QUERY on a serving endpoint. Experts can add example questions and guidelines in natural language to improve delegation.
MLflow Tracing records the inputs, outputs, latency, token usage and cost of every intermediate step in agent workflows that combine LLMs, retrievers, tools and sub-agents. Unity Gateway governs access to models, MCP servers and skills, and applies guardrails, rate limits and usage tracking.

What are common use cases?

The Supervisor Agent docs list market analysis across research reports and usage data, answering questions about internal processes while automating a ticket backlog, and speeding up customer service with policy, FAQ and account answers. Each case spans several specialized agents, so clear descriptions and scoped permissions matter.

FAQs

Do I need Supervisor Agent to build a multi-agent system on Databricks?

No. For new multi-agent systems in code, the docs point to the Agent Bricks CLI. For low-code agents, they point to Genie Agents.

Can I use LangGraph or another open-source framework?

Yes. Agent Bricks doesn't require a specific framework. The CLI ships templates for LangGraph and the OpenAI Agents SDK, and agentbricks init --framework langgraph --existing . prepares an existing LangGraph agent for migration.

Can I register a described agent in Unity Catalog?

Yes, as an agent service, which is in Beta. The comment field holds a description that shows in Catalog Explorer. Runtime invocation through a registered agent service isn't available during the Beta.

How do I debug a multi-agent run?

Open the run's trace in MLflow. Agent Runtime configures tracing to the experiment bound to your project, so each step is available for debugging.