Databricks · One-Pager · GenAI / Agents

Agent Bricks auto-optimized

A declarative way to build domain-specific AI agents on your governed data — you describe the task, Agent Bricks generates the evaluation and tunes the agent for you.

Beta Describe task → auto-eval → auto-optimize Verified June 2026 docs

In one line

Agent Bricks is a declarative, auto-optimized way to build and deploy domain-specific agents on Databricks: you pick a task type, point at your data, and it auto-generates an evaluation suite and tunes for quality and cost.

Why it matters

Building production agents normally means hand-crafting prompts, evals, and tuning loops. Agent Bricks collapses that: it uses Mosaic AI Research techniques to synthesize task-aware benchmarks and LLM judges, then optimizes the agent for you — so teams reach higher quality at lower cost without owning the agent-engineering plumbing.

Core concepts

Declarative build

Select a task type, write a natural-language description of the goal, and connect Unity Catalog data sources — no prompt or pipeline plumbing.

Auto-optimization

Automatically generates domain-specific synthetic data and task-aware benchmarks, then tunes the agent to trade off quality vs. cost.

Built-in evaluation

Creates evaluation suites and custom LLM judges per task, built on MLflow and Agent Evaluation; you steer quality via natural-language feedback and guidelines.

Four agent types

Information Extraction, Knowledge Assistant, Multi-Agent Supervisor, and Custom LLM — one prebuilt tile per common enterprise pattern.

On Mosaic AI

Sits on the Mosaic AI stack — Foundation Model APIs, Vector Search, Model Serving, MLflow — and deploys each agent as a governed serving endpoint.

UC governance

Data, functions, and connections come from Unity Catalog; the Supervisor enforces access controls so users only reach subagents they're authorized for.

How it works

Click a stage

1 · Pick agent type 2 · Point at data 3 · Auto eval + optimize 4 · Deploy endpoint

Pick agent type — choose Information Extraction, Knowledge Assistant, Multi-Agent Supervisor, or Custom LLM and describe the task in natural language.

Minimal example

Python · query a deployed Agent Bricks endpointfrom mlflow.deployments import get_deploy_client

client = get_deploy_client("databricks")
# each Agent Bricks tile is served as a chat-style endpoint
resp = client.predict(
    endpoint="ka-product-support",            # the deployed agent's serving endpoint
    inputs={"messages": [{"role": "user",
                          "content": "What is the refund window?"}]}
)
print(resp["choices"][0]["message"]["content"])

Gotchas & limits

Agent Bricks is in Beta; admins enable it (and per-tile features like the Custom LLM and Supervisor SDK/API) from the workspace Previews page.
Custom LLM is for text-transformation tasks (summarize, classify, generate) — it is not a general chat or tool-calling agent; the UI Custom LLM tile is being repositioned as the legacy path.
Quality is only as good as the guidance you give: examples, labels, and natural-language guidelines drive the auto-tuning and the synthetic eval set.
Source data must live in Unity Catalog (volumes, tables, Vector Search / AI Search indexes); the agent's service principal needs the right grants.

Interview soundbites

Say this out loud

"Agent Bricks is the declarative, auto-optimized layer on Mosaic AI — you describe the task and connect your data, and it generates the evals and tunes the agent for quality and cost."
"Four prebuilt types map to common patterns: Information Extraction, Knowledge Assistant, Multi-Agent Supervisor, and Custom LLM."
"The differentiator is auto-evaluation: Mosaic AI Research techniques synthesize task-aware benchmarks and LLM judges, so optimization isn't guesswork — and it's all governed by Unity Catalog."

References

Verified against docs.databricks.com · June 2026