Agent Bricks: Building Production AI Agents on the Lakehouse

The Fast Version

Databricks Agent Bricks is the platform’s agent-building layer: you describe a task, point it at governed data in Unity Catalog, and it builds, evaluates and deploys an agent as a serving endpoint. It launched in Beta in June 2025. Specific components, including Custom Agents and Document Intelligence, are now generally available, but the platform as a whole carries no single GA label.

• Cost: you pay for the serverless compute, model serving and tokens the agent consumes, so evaluation runs and traffic drive the bill.
• Effort: the build is fast; the slow part is curating source documents, writing guidelines and getting subject matter experts to label answers.
• Risk: an agent over ungoverned or contradictory data answers confidently and wrongly, which is a data problem Agent Bricks cannot fix for you.
• Fit: strongest when the knowledge already lives in the lakehouse and a named business owner will judge quality.

Most enterprise agent projects do not fail at the model. They fail at the data underneath it. A Cloudera and Harvard Business Review Analytic Services survey published in March 2026 found only 7% of enterprises say their data is completely ready for AI. An agent built on the other 93% inherits every inconsistency in the source.

The engineering bench is already stretched, because the same people who would clean that data also keep the existing pipelines running. So when a business leader asks for “an agent that answers policy questions by next quarter,” the team needs a way to build it without hand-writing retrieval, evaluation and serving from scratch.

That is the gap Agent Bricks targets. This guide explains what it is, where its status actually stands, how to build a first agent step by step, and which team profiles should use it now.

What Agent Bricks Is, and What It Is Not

Databricks introduced Agent Bricks at Data + AI Summit in June 2025 with a simple pitch: describe the task in plain language, connect your enterprise data, and let the platform handle evaluation and tuning that teams used to do by hand. Since then it has grown from an auto-optimisation tool into a broader agent platform.

The current documentation groups the work into three building blocks:

• Knowledge Assistant builds a domain chatbot over your documents, with answers grounded in the sources you choose.
• Supervisor Agent orchestrates other agents and tools: Genie agents, agent endpoints, Unity Catalog functions, MCP servers and custom agents.
• Custom agents let engineers author agents in Python with any library, including LangGraph, LangChain, OpenAI and LlamaIndex, then deploy and evaluate them on the same platform.

What it is not: a replacement for data engineering. Agent Bricks reads what Unity Catalog governs. If your policy documents sit in five SharePoint folders with three conflicting versions of the travel policy, the agent will retrieve all three.

Where the Status Actually Stands in September 2026

Status gets misquoted constantly, so be precise in any business case. Agent Bricks launched in Beta. In April 2026, Databricks positioned it as the governed enterprise agent platform and confirmed Custom Agents as generally available. The Data + AI Summit 2026 update in June labelled Document Intelligence, the SQL functions that parse documents, as GA.

Unity AI Gateway, the governance layer for model and agent traffic, has its own GA announcement dated 4 August 2026, with service policies and agent services still in Beta. That is a GA date for the gateway, not for Agent Bricks. Other named features, such as Agent Memory Service and Genie Ontology, carry no status label in the June post. Check the label in your workspace before you promise an SLA on any of them.

How to Build Your First Agent With Agent Bricks

This sequence assumes a Knowledge Assistant, the most common first project. The mechanics carry over to a Supervisor Agent later.

1. Name the decision the agent supports. “Answer HR policy questions” is too broad. “Tell a regional manager which leave policy applies to a contractor in Texas” gives you a testable scope and a business owner.
2. Curate the sources in Unity Catalog. The Knowledge Assistant documentation accepts files in a Unity Catalog volume (txt, pdf, md, ppt and pptx, doc and docx), a table with a file column, or an AI Search index. It allows up to 10 knowledge sources, skips files over 100 MB, and skips documents over 500 pages. Remove superseded versions before you load anything.
3. Check the prerequisites. You need Unity Catalog, serverless compute, Model Serving access, a serverless usage policy with a nonzero budget, and a supported region. A region gap stops the project on day one.
4. Write guidelines, not prompts. Guidelines are plain-language rules, such as “always cite the policy section” or “decline questions about individual salaries.” They apply as soon as you save them.
5. Run a labeling session with subject matter experts. Add the questions real users ask, let experts judge the answers, and feed their natural-language feedback back in. This is the step teams skip and the one that decides quality.
6. Measure with MLflow. Use MLflow tracing and evaluation to track answer quality on a fixed question set. Without a baseline you cannot tell whether a change helped.
7. Deploy to an endpoint and govern the traffic. The agent becomes a serving endpoint you query from AI Playground, Python or REST. Put it behind your gateway policies before anyone outside the pilot group uses it.

For the retrieval design underneath step 2, our guide to context engineering for AI agents covers chunking, metadata and what to leave out of the context window.

What Goes Wrong on Real Agent Builds

These problems compound in order. Fix the first and the others shrink.

1. Conflicting source documents. Three versions of one policy produce answers that are each correct against one version. The agent looks unreliable when the real defect is document control. Fix it in the volume, not in the prompt.
2. No accountable judge. Without a named expert labeling answers, quality becomes a matter of opinion in a steering meeting. Assign one owner per domain who signs off on the question set.
3. No fixed evaluation set. Teams tweak guidelines, try a new model, and cannot say whether anything improved. Freeze 50 to 100 real questions with expected answers before week two.
4. Cost surprises from evaluation loops. Every evaluation run consumes tokens and serverless compute. Tag the workload and review spend weekly, the same way you watch any other lakehouse job.
5. Treating the pilot as the product. A demo that impresses on ten questions fails on the eleventh. Most hallucination complaints trace back to missing context, which we cover in why AI agents hallucinate on company data.

Notice that none of these is a model problem. They are process problems, which is why a Lean Six Sigma approach fits: define the defect (a wrong answer), measure it against a fixed set, analyse the root cause in the data, improve the source, and control with ongoing evaluation. That is the same DMAIC loop that stabilises a manufacturing line.

Agent Bricks Compared With Building It Yourself

The trade is ownership against speed. If your data already lives in the lakehouse, the governance and evaluation plumbing is the expensive part to rebuild, and Agent Bricks gives it to you.

If Your Team Looks Like This, Do This

Lakehouse in place, no agent yet

Your gold layer is governed in Unity Catalog and a business owner has a document-heavy question set. Start with one Knowledge Assistant in one domain, run the seven steps above, and hold the go-live decision until your evaluation set passes an agreed threshold.

Data still scattered across SharePoint and file shares

Do not build an agent yet. Spend the first month landing and de-duplicating the documents into Unity Catalog volumes with clear ownership. The agent build then takes days instead of a quarter of rework.

Engineers already running LangGraph agents elsewhere

Move the orchestration code onto custom agents so you keep your framework and gain MLflow tracing, evaluation and governed serving. Rewrite nothing that already works; migrate the plumbing around it.

Several single-purpose agents and a Genie space

You are ready for a Supervisor Agent that routes questions to the right agent or Genie space. Define routing rules and ownership per sub-agent first, or failures become impossible to trace. If you want help sequencing that build, our Databricks consulting team runs it with AIM-IT, and our AI agent implementation services page explains how we scope the work.

Frequently Asked Questions (FAQs)

Is Databricks Agent Bricks generally available?

Not as a single platform-wide label. Agent Bricks launched in Beta in June 2025. Databricks has since confirmed specific components as GA, including Custom Agents and Document Intelligence, while other features carry no stated status. Check each feature’s label before you commit to production terms.

Can I use my own framework with Agent Bricks?

Yes. Custom agents can be written in Python with any authoring library, including LangGraph, LangChain, OpenAI and LlamaIndex. You keep the framework and gain Databricks tracing, evaluation and serving around it.

What data can a Knowledge Assistant read?

Files in a Unity Catalog volume, a Unity Catalog table with a file column, or an AI Search index. Supported file types include txt, pdf, md, PowerPoint and Word documents, with limits on file size and page count.

Has anyone shipped Agent Bricks agents to production?

Databricks reports broad usage, but independent failure reports are still scarce, so treat launch coverage as marketing until your own evaluation set says otherwise. The teams that ship are the ones with curated sources, a named expert judge and a frozen evaluation set.