The Mosaic AI Agent Framework is the code path for agents on Databricks: author the agent in any framework, wrap it in MLflow's ResponsesAgent interface, trace it and deploy it to Model Serving. Agent Bricks is the no-code path: Knowledge Assistant, Multiagent Supervisor and Information Extraction. The Databricks Generative AI Engineer Associate exam tests both, and when to pick which.
Last updated October 2026.
You build the loop in LangGraph, the OpenAI SDK or plain Python, then subclass MLflow's ResponsesAgent and call your agent inside its predict method. That one wrapper gives the agent a standard request and response contract and an inferred signature, which AI Playground, mlflow.genai.evaluate() and Agent Monitoring are all built against (Chapter 25).
The exam names four more pieces:
mlflow.genai.evaluate() runs judges and custom Scorers over those traces, so a scorer can grade a retrieval step or a tool result, not only the final answer (Chapter 50, Chapter 49).agents.deploy() call creates a Model Serving endpoint. Inference tables, Agent Monitoring and Mosaic AI Gateway then track it in production (Chapter 53, Chapter 54).Agent Bricks are prebuilt, declarative agents: you configure a use case with data, instructions and examples through a UI, and Databricks builds, auto-tunes and evaluates the agent behind it with LLM judges and the feedback you provide. The result is a governed endpoint with MLflow Tracing. The guide names three, and each solves one problem shape (Chapter 7).
The objective reads "determine how and when to use Agent Bricks", and knowing when not to is half of it. Reach for a brick when the problem matches one of the three shapes: you trade code-level control for speed, built-in evaluation and auto-optimization. Drop to the Agent Framework when you need custom orchestration logic, unusual output contracts, or a framework your team already owns. The two also mix: a Multiagent Supervisor can route to a custom Agent Framework endpoint alongside its bricks.
| Mosaic AI Agent Framework | Agent Bricks | |
|---|---|---|
| You supply | Agent code in any framework, wrapped in ResponsesAgent | Data, instructions and examples through a UI |
| Orchestration | Yours: tools, routing and prompts in code | Databricks builds and auto-tunes it |
| Evaluation | mlflow.genai.evaluate() with judges and custom Scorers over traces | Built in, with LLM judges plus your SME feedback |
| Deployment | Register to Models in Unity Catalog, then agents.deploy() | An endpoint, created for you |
| Reach for it when | Custom logic, unusual output contracts, an existing framework | The problem is Q&A over documents, routing across specialists, or documents to columns |
Comparison as taught in Chapters 7 and 25 of the Certified course, from the official exam guide's objectives and the Databricks documentation.
When a question is about what your numbers say rather than your documents, the agent needs a tables specialist, and on Databricks that is a Genie Space: a curated natural-language interface over selected Unity Catalog tables. Code reaches it through the asynchronous Conversation API (start a conversation, poll to COMPLETED, read the rows), wrapped as a GenieAgent subagent inside an Agent Framework agent, or with no code as a Genie Space specialist behind the Multiagent Supervisor. Unity Catalog still executes the SQL, so the agent's reach ends where its table permissions end (Chapter 26).
MCP servers are how agents reach tools over the Model Context Protocol, and the guide names three kinds. Managed servers are hosted by Databricks and already expose Unity Catalog functions, Vector Search, Genie and SQL, so they come first. Custom servers are yours, typically hosted as a Databricks App, for proprietary logic nothing managed covers. External servers are third-party services reached through governed OAuth. One of the guide's two select-two sample questions lives here (Chapter 39).
Agents run through four of the six sections. Section 1 (Design Applications) asks you to define and order tools for multi-stage reasoning and to determine when to use Agent Bricks. Section 3 (Application Development) has "utilize MLflow and Agent Framework for developing agentic systems" and "enable multi-agent systems to leverage Genie Spaces or the conversational API". Section 4 (Assembling and Deploying Applications) adds integrating managed, external and custom MCP servers, testing individual components of an agent in CI/CD (Chapter 41), and building a user-facing interface in Databricks Apps, Slack or Teams, where the app backend calls the endpoint and a token never sits in browser JavaScript (Chapter 42).
Section 6 (Evaluation and Monitoring) closes the loop: evaluate agent performance with MLflow scoring and tracing, use custom Scorers, use inference tables and Agent Monitoring on a live endpoint, and use AI Gateway to track an agent deployed through the Agent Framework. That is a dozen of the 56 objectives, so the framework-versus-brick judgment is worth getting right.
The Mosaic AI Agent Framework is the Databricks path for building agents in code. You author the agent in any framework, wrap it in MLflow's ResponsesAgent interface, trace it with MLflow Tracing, evaluate it with mlflow.genai.evaluate(), register it to Models in Unity Catalog and deploy it to Model Serving with agents.deploy().
Knowledge Assistant, Multiagent Supervisor and Information Extraction. Knowledge Assistant answers questions over your documents with citations, Multiagent Supervisor routes one request across specialist agents and composes a single answer, and Information Extraction turns a corpus of documents into a structured Delta table from a schema you describe.
Use an Agent Brick when the problem matches one of its three shapes: question answering over documents, routing across specialists, or documents to columns. Drop to the Agent Framework when you need custom orchestration logic, an unusual output contract or a framework your team already uses, and remember a Multiagent Supervisor can route to a custom agent endpoint too.
A Genie Space is the structured-data specialist: a curated natural-language interface over selected Unity Catalog tables. An Agent Framework agent reaches it through the Conversation API as a GenieAgent subagent, and a Multiagent Supervisor can add a Genie Space as a specialist with no code, so one front door answers both document and table questions.
Managed MCP servers are hosted by Databricks and expose Unity Catalog functions, Vector Search, Genie and SQL with no server to run. Custom servers are ones you build and host, typically as a Databricks App, for proprietary logic. External servers are third-party services that speak MCP, reached through governed OAuth, and all three are governed through Unity Catalog.
Chapter 7 (Agent Bricks) sits in the free Unit 1 of Certified's Generative AI Engineer Associate course, and Chapters 25, 26, 39 and 50 carry the Agent Framework, Genie Spaces, MCP servers and agent evaluation. Units 1 and 2 are free to start.
Open Chapter 7 →