Week of Agents | New in Agent Bricks: MCP, Document Intelligence & More

Summary

• Agent Bricks Expansion: Agent Bricks now includes MCP support via a new MCP Catalog and Marketplace for interoperability across models and frameworks.
• Document Intelligence: New ai_parse_document function intelligently extracts text, tables, and figures from PDFs with AI-generated descriptions and bounding boxes.
• Unified Governance: Information extraction, knowledge assistants, and agents all operate within Databricks Data and AI platform governance and evaluation frameworks.

Week of Agents | New in Agent Bricks: MCP, Document Intelligence & More

This week, Databricks is kicking off another “Week of Agents,” unveiling new Agent Bricks capabilities that make it easier to build trusted, production-ready AI agents. The latest updates expand governance with the new MCP Catalog and Marketplace, introduce intelligent document processing with ai_parse_document, and advance interoperability across any model, framework, or MCP server.
Agent Bricks now gives every team the flexibility to build, evaluate, and scale AI agents confidently—on one open, governed platform. Tune in throughout the week for deeper dives into these releases!

Chapters

• 0:00 Introduction to Agent Bricks
• 0:39 Unstructured data extraction
• 1:43 Knowledge assistant brick
• 2:38 Genie text-to-SQL agent
• 3:32 Multi-agent orchestration
• 4:25 Production deployment and conclusion

FAQs

What is the MCP Catalog and Marketplace?

This video announces new MCP (Model Context Protocol) support through a catalog and marketplace, enabling Agent Bricks to work with any LLM model, framework, or MCP server.

How does ai_parse_document extract information?

The ai_parse_document function automatically extracts every element from documents including text, tables, and figures. It generates AI descriptions and bounding boxes, making unstructured documents queryable.

How can I handle unstructured data like PDFs?

This video demonstrates using the information extraction brick with ai_parse_document to automatically identify key fields and entities, fine-tune extraction, and evaluate quality continuously.

Can I improve agent accuracy over time?

Yes. The knowledge assistant brick lets you label sessions, provide feedback on responses, and add guidelines for accuracy and tone. Subject matter experts can guide agents without ML engineering expertise.

Full transcript

0:00 Introduction to Agent Bricks

Hi everyone, this week we're excited to announce some new releases around Agent Bricks. Throughout this week, we'll be showcasing expanded MCP support, powerful new intelligent document processing functionality, and expanded model support for all leading providers. Ahead of these announcements, we want to recap on what Agent Bricks is and how it's already helping users build AI agents that reason accurately on your data while delivering domain specific quality, continuous evaluation, all with robust governance.
Let's walk through a few common challenges our users face and how agent bricks is already helping address them. So, one of the most common challenges we see is working with unstructured data, things like PDFs, contracts, or scanned reports. These documents often contain critical information, but that data isn't structured or ready for analysis. Our information extraction brick helps you solve exactly that. It lets you automatically identify key fields and entities, fine-tune extraction for your data, and continuously evaluate quality and cost.

0:39 Unstructured data extraction

Using the power of our new AI parse document function, you can now extract every element from their documents, text, tables, and even figures with AI generated descriptions and bounding boxes. The result, unstructured documents become queryable and governed just like any other table in your lakehouse. No brittle pipelines, just one unified way to turn unstructured content into data that you can use for analytics, retrieval or other agents.
Now, another common challenge we see is that teams want to interact with their information, whether it's internal documentation, reports, or knowledge bases without needing to build complex retrieval pipelines. And that is where our knowledge assistant brick comes in. This brick lets you spin up an agent that can answer questions about your data using natural language powered by an LLM in the background. But it's not just about getting answers.

1:43 Knowledge assistant brick

It's about being able to steer the agents quality and tone. So you can start labeling sessions, provide feedback on responses, and add guidelines for accuracy, tone, and reasoning style. That feedback loops directly into the agent, meaning that your subject matter experts and not just your machine learning engineers can guide and improve the system over time. This creates a fast, intuitive way to reason over your domain data while keeping humans firmly in the loop.

2:38 Genie text-to-SQL agent

Now, of course, most organizations don't just deal with unstructured information. They also rely heavily on structured sources like databases and tables. That's where Genie comes in. So, Genie is our text to SQL agent that lets you query your structured data conversationally using natural language. It's perfect for analysts and business users and really any user that wants quick insights without having to write bespoke SQL.
Just like the other bricks, Genie supports iterative tuning. So, you can add specific SQL instructions or domain specific rules and jargon and evaluation metrics to continuously improve the accuracy and reliability of your results. With knowledge assistant and Genie together, you now have highquality natural language access to both unstructured and structured data. All governed, consistent, and customizable. Now, once your data is structured and governed, the next step is putting it to work.
That's where agent bricks multi- aent supervisor comes in. It acts as an intelligent orchestrator that ties multiple agents together, decides which one to use for a given query, manages the back and forth between them, and returns a unified highquality answer, complete with references for transparency. And now, we're excited to announce native MCP support in the multi-agent supervisor. With MCP, your agents can do even more. securely accessing governed thirdparty data from your catalog or the marketplace and triggering actions in your system like running an analysis, creating a report, or maybe filing a ticket.

3:32 Multi-agent orchestration

And just like the other bricks, you can still run labeling sessions here, letting your domain experts steer and evaluate multi-agent reasoning and quality over time. So, that's a quick look at how Agent Bricks is helping build intelligent domain specific agents that reason accurately on your data. All with built-in evaluation, optimization, and governance. This week, you'll see new exciting updates and releases across all of these bricks along with powerful new capabilities that make it even easier to build and deploy production grade AI systems.

4:25 Production deployment and conclusion

We can't wait to show you what's next. So, thanks for tuning in and uh enjoy the week of agents.

Learn more about the Databricks Data and AI platform.