Databricks Agent Bricks
Also known as: Databricks Mosaic AI, Mosaic AI, MosaicML, Mosaic AI Agent Framework
Databricks' agent platform, now sold as Agent Bricks, for building, governing and monitoring multi-agent systems on the Databricks Data Intelligence Platform.
Mosaic AI is the name Databricks used for the artificial-intelligence layer of its Data Intelligence Platform, which originated with the 2023 acquisition of MosaicML. Databricks now sells the agent part of that layer as Agent Bricks, described as the control plane for governing, managing, and monitoring AI agents across an enterprise, and the Mosaic AI name no longer appears on its product pages or documentation.
Teams build agents two ways. Managed agents are configured rather than coded: Knowledge Assistant answers questions over a company's documents, Supervisor Agent coordinates other agents, Unity Catalog functions, MCP servers, and code execution into one system, and Genie Agents answer natural-language questions over governed data. Code-first agents are built with the Agent Framework in Python using common agent libraries. Agents draw on AI Search (formerly Vector Search) for retrieval, can keep long-term memory per user or team in Lakebase (in beta), and are served as REST endpoints, deployed as Databricks Apps, or scheduled on recurring workflows.
Governance is the core selling point. Unity Catalog applies role-based access to models, tools, and connections with lineage, end users reach only the agents and data they are granted, and service policies can block sensitive data, prompt injection, and unsafe content or hold a risky tool call for human approval. MLflow traces every interaction, tool call, and model call, and the same scorers and LLM judges used to compare versions run on a sample of production traffic. Admins can connect OpenAI, Anthropic, Bedrock, Gemini, and other providers with their own keys through Unity Gateway alongside hosted models.
Pricing follows Databricks consumption billing in Databricks Units on AWS, Azure, and Google Cloud, with a free trial and committed-use discounts. It is most compelling for organizations whose data already lives on Databricks and want agents built and governed in the same place.
Vendor details
Canonical URL
https://www.databricks.com/product/artificial-intelligence/agent-bricks
Category
Multi-agent platform
Subcategory
Enterprise agent platform on Databricks (Mosaic AI)
Funding status
Agent Bricks is the agent layer of Databricks, the company used by 15,000+ organizations. It is built on Databricks' 2023 acquisition of MosaicML, sold for a time as Mosaic AI, now fully integrated across the Databricks Data Intelligence Platform.
Company status
first party product
Use cases & customers
Primary use cases
Target customers
Deployment options
Integrations
Native to the Databricks Data Intelligence Platform: connects enterprise data in Delta Lake to LLMs via AI Search (formerly Vector Search), Model Serving, Foundation Model APIs (serving OpenAI, Anthropic, and others), MLflow tracing/evaluation, and Unity Catalog governance. Supports tool-calling agents, RAG, and multi-agent systems authored in LangChain, LangGraph, or native Python, with a Tool Catalog for function-calling and an AI Gateway governing LLMs and MCP. Runs on AWS, Azure, and GCP.
In practice
Your agents need enterprise data, but the data and the models live in different worlds. Mosaic AI builds agents on Databricks where the data, models, and governance sit in one environment under Unity Catalog.
You can't tell whether your agent's output is actually good. Mosaic AI's Agent Evaluation grades quality with built-in AI judges and human review before you ship.
Tuning an agent for cost and quality is endless manual work. Databricks' Agent Bricks auto-generates task benchmarks and optimizes the agent against them.
Sources & related URLs
Research sources
Agentic Index coverage score
13.0 / 14 capabilities · 93%
| Integrations & Tool Calling | Full |
|---|---|
|
Agents take tools from managed, external and custom MCP servers, Unity Catalog functions, Databricks data and a sandboxed code execution tool, with role based access applied to models, tools and connections. Sourcedocs.databricks.com/agents/mcp-toolsread 2026-09-27 |
|
| Workflow Orchestration | Full |
|
Supervisor Agent delegates work across Genie Agents, agent endpoints, Knowledge Assistant endpoints, Unity Catalog functions, MCP servers and custom agents and synthesizes the results, and code first multi agent systems are built with the Agent Framework. Sourcedocs.databricks.com/aws/en/agents/agent-bricks/multi-agent-supervisorread 2026-09-27 |
|
| Knowledge Grounding & RAG | Full |
|
AI Search (formerly Vector Search) is a governed vector index built into the platform that agents query for retrieval, and Knowledge Assistant builds a question answering agent over the customer's documents. Sourcedocs.databricks.com/ai-search/ai-searchread 2026-09-27 |
|
| Human Oversight & Guardrails | Full |
|
Service policies on MCP and model services can return an "ASK" decision that holds an interaction for human approval before it proceeds, such as a destructive MCP tool call, and also block PII, prompt injection and unsafe content; Supervisor Agent asks users to approve web search queries. Service policies are in beta and do not yet cover agent services. Sourcedocs.databricks.com/data-governance/unity-catalog/service-policiesread 2026-09-27 |
|
| Security, Identity & Governance | Full |
|
Unity Catalog applies role based access to models, tools and connections with lineage, end users reach only the subagents and data they are granted, and model provider credentials stay behind privileges; the compliance docs name profiles for HIPAA, IRAP, PCI-DSS and FedRAMP High and Moderate, with reports through the Databricks Security and Trust Center. Sourcedatabricks.com/product/artificial-intelligence/agent-bricksread 2026-09-27 |
|
| Observability & Auditability | Full |
|
MLflow Tracing captures every interaction, tool call and model invocation of an agent with no code changes, each session shareable by link, and Unity Gateway's unified tracing adds tool and skill use to model requests under Govern > Traces. Sourcedocs.databricks.com/mlflow3/genai/tracing/overviewread 2026-09-27 |
|
| Memory & State Persistence | Full |
|
Managed agent memory (Beta, 16 Sep 2026) stores long term memory that persists across conversations in Lakebase, scoped per verified end user or to a team, project or organization key, with semantic search and the source session recorded; managed sessions keep durable conversation history. It works with agents on any framework. The memory page documents listing and searching entries but not deleting one user's entries. Sourcedocs.databricks.com/aws/en/agents/agent-memory/managed-memoryread 2026-09-27 |
|
| Deployment & Data Residency | Full |
|
The platform runs on AWS, Azure and Google Cloud with a published supported regions list, can be configured against the customer's own cloud account, and deploys agents to Model Serving endpoints or serverless Databricks Apps. Sourcedocs.databricks.com/resources/supported-regionsread 2026-09-27 |
|
| Prebuilt Agents, Templates & Packs | Full |
|
Named managed agents ship ready to configure: Knowledge Assistant answers questions over the customer's documents, Supervisor Agent coordinates other agents and tools, and Genie Agents answer natural language questions over governed data; each does its own job without the others. Sourcedocs.databricks.com/agentsread 2026-09-27 |
|
| Triggers & Channel Coverage | Full |
|
Agents are served as REST endpoints or scheduled on recurring workflows through Lakeflow Jobs, which run without a person starting them, and Genie reaches users in Slack and on mobile. The scheduled workflow is what lets work reach the agent unprompted. Sourcedatabricks.com/product/artificial-intelligence/agent-bricksread 2026-09-27 |
|
| Model Flexibility & Routing | Full |
|
Admins connect external providers with their own keys, OpenAI, Azure OpenAI, Anthropic, Amazon Bedrock, Microsoft Foundry, Google Gemini Enterprise or a custom provider, as Unity Catalog securables granted per user, alongside hosted DeepSeek, Llama, Gemini, Claude and GPT models; September 2026 notes add Smart Routing defaults set centrally. Sourcedocs.databricks.com/ai-gateway/model-provider-servicesread 2026-09-27 |
|
| APIs, SDKs & MCP Extensibility | Full |
|
Agents are served as REST APIs, Supervisor Agents can be created programmatically with the Databricks SDK and queried through the API, Declarative Automation Bundles manage resources as code, and the Genie One MCP server (GA 25 Sep 2026) exposes Genie to outside MCP clients under Unity Catalog permissions. Sourcedocs.databricks.com/aws/en/agents/agent-bricks/multi-agent-supervisorread 2026-09-27 |
|
| Testing, Debugging & Optimization | Full |
|
MLflow scorers, LLM judges and code checks, grade traces from the UI or across a dataset with mlflow.genai.evaluate() to compare app versions, the same scorers run on a sample of production traffic, and Agent Bricks generates task specific synthetic data to optimize quality. Sourcedocs.databricks.com/mlflow3/genai/eval-monitorread 2026-09-27 |
|
| Browser & Computer Use | Not documented |
|
Supervisor Agent carries a sandboxed code execution tool with no internet access and a web search tool, and agents work through APIs, data and MCP; no browser or computer use capability is documented. Sourcedocs.databricks.com/aws/en/agents/agent-bricks/multi-agent-supervisorread 2026-09-27 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Recent platform changes
Databricks released the Agent Bricks CLI in beta, which scaffolds a custom code agent from a LangGraph or OpenAI Agents SDK template, runs it locally and deploys it to the Databricks agent runtime. One config file declares the agent's tools, memory, sessions and tracing, and deployment provisions them and grants the agent's own identity access.
Bears on: Workflow orchestration
View sourceUnity Gateway adds centralized configuration for coding agents, including approved models, MCP servers, skills, and routing settings. Its new ug CLI applies those settings and handles authentication when launching agents such as Claude Code and Codex. Unified tracing now captures the tools and skills used alongside model requests.
Bears on: Observability / auditability
View sourceDatabricks announced Omnigent, an open-source meta-harness above existing agents that adds multi-agent composition, contextual policies, real-time collaboration, cloud execution, and multi-harness authoring.
Bears on: Workflow orchestration
View sourcePricing
Free trial · usage based (DBUs)
usage (DBUs)
Included quota
There is no fixed seat price. A free trial lets teams evaluate the platform, after which usage is billed by DBU consumption across model serving, vector search, agent execution, and jobs, plus per token Foundation Model API usage. Enterprise committed use contracts bundle volume at a discount. Available on AWS, Azure, and GCP.
What is public
Public: the consumption based model, per DBU rates by cloud and workload, and per token Foundation Model API rates. Not public as a single figure: total cost, which depends entirely on consumption, and enterprise committed use discounts, which are negotiated.
Billing mechanics
Consumption billing in DBUs for compute and DSUs for storage at per second granularity, with each workload (model serving, AI Search, agent execution, jobs) metered at price list rates that vary by cloud and product. Model APIs bill per token or by provisioned throughput. Committed use contracts give discounts. When Databricks runs against the customer's own cloud account, the cloud provider bills its resources separately. Non Azure accounts can pay monthly by card and cancel anytime.
Cost watchouts
DBU consumption spans multiple workloads (model serving, vector search, agent execution, jobs) that bill in parallel, and underlying cloud compute is billed by the cloud provider on top. Foundation Model API token costs scale with usage. Real spend is hard to predict from the rate card alone and depends on architecture and volume.
Variable cost rationale
Billing is purely consumption based on DBUs and per token model usage, so cost scales directly with agent activity, model serving, vector search, and job volume. Long running or high throughput agent systems accumulate cost across several workloads at once, and underlying cloud infrastructure is billed separately, so exposure is high, though committed use contracts and provisioned throughput can make spend more predictable.
Additional watchouts
Consumption based billing means cost tracks usage across several workloads plus underlying cloud compute, which is billed separately. Committed use pricing is negotiated. Realistic budgeting requires estimating DBU and token consumption for your workloads.
Overage / add-ons
There are no fixed caps; usage bills by DBU consumption and per token model usage. Committed use contracts prepay volume at a discount, and provisioned throughput reserves capacity for steadier cost.
Sales call required
Mixed (some tiers require a call)
Free / trial
Free trial of the Databricks platform; some trials carry credits that Databricks sets after the account is created.
Lowest paid plan
Pay as you go consumption on Databricks Units, with published per DBU and per token rates, above a free trial.
Commercial notes
The agent layer of the Databricks Data Intelligence Platform, sold under the Agent Bricks name, so pricing follows the Databricks consumption model rather than a standalone agent subscription. Most compelling for teams already on Databricks, where data, models, governance and agents share Unity Catalog.
Key ambiguities
Total cost depends entirely on consumption across workloads and on the cloud, so a single entry price is not meaningful. Enterprise committed use discounts are negotiated and not public.
Support SLA / resale
Support tiers follow the Databricks platform (business, enhanced, production, and mission critical), with enterprise SLAs available on committed contracts. Available through AWS, Azure, and GCP marketplaces.
Missing data
Per DBU and per token rates vary by cloud, region and workload, so no single figure applies; enterprise committed use discount levels are negotiated and not public.
Related vendors
- CrewAI — Open source multi agent framework (Python, MIT) of role based Crews…
- Relevance AI — Platform for building and deploying a workforce of AI agents with a…
- Agent Zero — Open-source multi-agent framework that gives its agent a full Linux…
- Akka — Enterprise agentic AI platform that builds and runs multi agent…
- Artian — Enterprise multi agent AI platform for financial services: a Planner…
- Dataiku — Enterprise AI and machine learning platform with an agentic layer…
Alternatives to Databricks Agent Bricks
The closest documented capability profiles to Databricks Agent Bricks among multi-agent platforms tracked by Agentic Index, ordered by similarity on the same 14 point evidence the rankings use. No vendor pays for placement.
- Dataiku13.0 / 14Matches Databricks Agent Bricks across all 14 documented capabilitiesDatabricks Agent Bricks vs Dataiku →
- Relevance AI13.0 / 14Matches Databricks Agent Bricks across all 14 documented capabilities
- Akka12.5 / 14A lighter documented profile than Databricks Agent BricksDatabricks Agent Bricks vs Akka →
- CrewAI13.0 / 14Adds documented Browser & Computer UseDatabricks Agent Bricks vs CrewAI →
- Legion Intelligence11.5 / 14A lighter documented profile than Databricks Agent Bricks
- Rasa11.5 / 14A lighter documented profile than Databricks Agent Bricks
Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded