Agent Frameworks, MLflow and Genie in Multi Agent Systems

What it is

An agent is an application in which the LLM decides which tools to call, such as a retriever, a Unity Catalog function or a Genie space, and in which order. Open source libraries such as LangChain and LangGraph supply the building blocks.

Mosaic AI Agent Framework and MLflow 3 turn that code into something you can trace, evaluate, log to Unity Catalog and deploy. A multi agent system adds a supervisor that routes each question to a specialist, for example a Genie space for questions over tables.

How it works

Choosing LangChain or a similar tool

The databricks_langchain package connects LangChain to Databricks, for example ChatDatabricks for serving endpoints. MLflow traces these libraries through autologging, such as mlflow.langchain.autolog().

Choose Agent Bricks when the use case matches one of its patterns and speed matters. Choose a custom Agent Framework agent when you need your own logic, tools or framework.

Building with Agent Framework and MLflow

Databricks recommends wrapping a custom agent in the MLflow ResponsesAgent interface, which gives a standard chat request and response schema that works with the Review App, AI Playground and serving. The older ChatAgent interface still exists, but ResponsesAgent is the current recommendation.

The usual steps are:

import mlflow
mlflow.set_registry_uri("databricks-uc")
info = mlflow.pyfunc.log_model(
    name="agent", python_model="agent.py",
    resources=resources,
    registered_model_name="main.support.support_agent")

The agent file ends with mlflow.models.set_model(agent), so MLflow logs the code itself (models from code). The resources list names the endpoints, indexes, functions and Genie spaces the agent uses, so the deployed agent can authenticate to them.

Tools come from Unity Catalog functions, whose COMMENT text becomes the tool description the LLM reads, or from MCP servers: managed MCP servers for Vector Search, Genie spaces and Unity Catalog functions, external MCP servers through connections, and custom MCP servers hosted as Databricks Apps.

Evaluation versus monitoring

The same scorers can run in both phases, which keeps the quality definition consistent. Bad production traces are added to the evaluation set, so each new version is tested on real failures.

Genie spaces in multi agent systems

A Genie space (shown as a Genie Agent in newer docs) turns natural language into SQL over a curated set of Unity Catalog tables. Agents use it in three ways:

Genie runs queries with the caller's Unity Catalog permissions, so the user or service principal needs access to the space and SELECT on its tables. Conversation context is kept by passing the returned conversation ID on the next call. For a no code setup, Agent Bricks Multiagent Supervisor (Supervisor Agent in newer docs, which mark it as legacy) coordinates Genie spaces with other agents.

How the exam asks about this

• "A supervisor must answer sales figures from governed tables and policy questions from documents." Add a Genie space as the structured data worker and a retriever agent for documents.
• "Find which tool call made a slow agent slow." Use MLflow Tracing.
• "Detect quality drift in production where no reference answers exist." Run monitoring with judges that need no ground truth.