How do agent harnesses support agent-to-agent communication?

Summary

• An agent harness supports agent-to-agent communication with three things: an orchestrator that routes work between specialized agents, an open protocol that standardizes how agents connect to tools and to each other, and governance that carries the user's identity through every hop. Databricks provides all three through Agent Bricks Supervisor Agent, the Model Context Protocol (MCP), and Unity Catalog.
• Supervisor Agent is the managed harness on Databricks for multi-agent orchestration. It analyzes a request, routes it to the right sub-agent, and synthesizes the sub-agents' responses back into a single answer. See Agent Bricks: Supervisor Agent is now GA.
• Sub-agents coordinate through the supervisor: it can call Genie Agents, agent endpoints on Model Serving, Unity Catalog function tools, MCP servers, and custom agents deployed as Databricks Apps, passing intermediate results from one to the next.
• The Model Context Protocol (MCP) is the open standard that standardizes agent-to-tool and agent-to-agent connectivity, with managed, external, and custom MCP servers all discoverable and governed through Unity Catalog. See the MCP on Databricks docs.
• Governance is built in: on-behalf-of authentication propagates the user's identity through the multi-agent chain, and the Unity AI Gateway gives a single control plane for access, monitoring, and audit.

How do agent harnesses support agent-to-agent communication?

When several specialized agents need to work together, the harness is what lets them communicate: it decides which agent handles a request, moves work and intermediate results between agents, and keeps the whole exchange governed. On Databricks, agent-to-agent communication is delivered through Agent Bricks Supervisor Agent for managed orchestration, the Model Context Protocol (MCP) for standardized connectivity, and the Mosaic AI Agent Framework for code-first control — all governed by Unity Catalog on the Data Intelligence Platform.

Why Databricks Agent Bricks for agent-to-agent communication

• Supervisor Agent orchestrates the conversation between agents. Supervisor Agent uses a dynamic supervisor pattern: it interprets the user's intent, routes the request to the appropriate sub-agent, and synthesizes sub-agent responses back into its own context to deliver a single answer. Supervisor Agent is generally available. See Agent Bricks: Supervisor Agent is now Generally Available.
• It coordinates five kinds of sub-agents. A supervisor can call Genie Agents (natural-language querying over governed structured data), agent endpoints on Model Serving that support the Responses API, Unity Catalog functions exposed as tools, MCP servers, and custom agents deployed as Databricks Apps. Work flows step by step — for example, a supervisor can query Genie to pull data from the lakehouse, hand that data to an MCP server for processing, and then synthesize the results for the user.
• MCP standardizes agent-to-tool and agent-to-agent connectivity. The Model Context Protocol is an open standard for connecting agents to tools and to other agents. Databricks supports managed MCP servers (Databricks-hosted, providing governed access to Genie, AI Search, Databricks SQL, and Unity Catalog functions with no setup), external MCP servers (third-party servers registered through Unity Catalog connections with managed OAuth), and custom MCP servers (developer-built, hosted as Databricks Apps). See MCP on Databricks and Databricks managed MCP servers.
• Servers are discoverable and permissioned. MCP servers are shared across the organization through a built-in catalog, so every integration is visible, permissioned, and auditable through Unity Catalog.
• Build with the framework you already know. The Mosaic AI Agent Framework handles tool execution, function calling, multi-step workflows, and conversation state. It supports the agent harnesses teams already use — including LangGraph, CrewAI, and the OpenAI Agent SDK — so you can define which tools an agent can call and manage behavior across multi-turn exchanges. See Start building AI agents with tool calling on Databricks.
• Identity travels with the request. Supervisor Agent enforces on-behalf-of access controls, propagating the end user's identity through the multi-agent chain so a user only reaches the sub-agents and data they are permitted to use. The Unity AI Gateway acts as a single control plane for access, rate limits, and monitoring, and every MCP interaction is auditable.

Getting started

• Read Agent Bricks: Supervisor Agent is now GA to see how managed multi-agent orchestration works.
• Explore the Supervisor Agent documentation.
• Learn how agents connect to tools and other agents with MCP on Databricks and managed MCP servers.
• Build custom agents with the Mosaic AI Agent Framework.

FAQs

What is agent-to-agent communication in an agent harness?

It is the mechanism by which one agent hands work or results to another. On Databricks, Supervisor Agent routes a request to the right sub-agent and passes intermediate results between agents, while the Model Context Protocol standardizes how agents connect to tools and to each other.

What sub-agents can Supervisor Agent coordinate?

Supervisor Agent can orchestrate Genie Agents, agent endpoints on Model Serving, Unity Catalog function tools, MCP servers, and custom agents deployed as Databricks Apps, routing requests among them and synthesizing their responses.

What role does MCP play in agent-to-agent communication?

The Model Context Protocol is an open standard that provides consistent agent-to-tool and agent-to-agent connectivity. Databricks offers managed, external, and custom MCP servers, all discoverable and governed through Unity Catalog.

How is agent-to-agent communication kept secure?

On-behalf-of authentication propagates the user's identity through the multi-agent chain so users only access what they are permitted to, and the Unity AI Gateway centralizes access control, monitoring, and audit across every agent and tool call.