Databricks now offers a full stack for building AI agents on lakehouse data: Agent Bricks for building and evaluating agents, Genie for natural-language analytics, Unity Catalog for governing data, tools and agents, and Unity AI Gateway for controlling model and tool access. The older LakehouseIQ name is history; its idea of a semantic layer that understands your business now lives in Genie and Databricks Assistant.
Updated October 2026
What changed in 2026
- At Data + AI Summit 2026, Databricks expanded Agent Bricks into what it calls a comprehensive agent platform, with frontier and open models, support for harnesses such as LangGraph, CrewAI, the Claude Code SDK and OpenAI Agent SDKs, and deployment to Databricks Apps.
- Context features arrived: MCP support in Unity Catalog for external sources, Genie Ontology for learning business semantics, an agent memory service powered by Lakebase, and Document Intelligence SQL functions (ai_parse_document, ai_extract, ai_classify) announced as generally available.
- Control features arrived: Unity AI Gateway for access, budgets and routing across agents, models, MCP servers and skills, plus agent traces and monitoring and SQL-based contextual policies.
- Enterprises are scaling agents, but carefully. McKinsey's State of AI 2026 reports 40% of $1B+ revenue companies scaling AI agents, while Gartner expects over 40% of agentic AI projects to be canceled by 2027.
From LakehouseIQ to Agent Bricks: a short history
| Year | Milestone |
|---|---|
| 2020–2022 | The lakehouse pattern (Delta Lake plus a SQL warehouse) becomes Databricks' core; ML runs through MLflow. |
| 2023 | LakehouseIQ is announced as a knowledge engine that learns an organisation's data, usage and jargon to power natural-language features. |
| 2024–2025 | Mosaic AI tools for agents and evaluation; AI/BI Genie for natural-language analytics; Agent Bricks introduced. |
| 2026 | Agent Bricks broadens into a full agent platform with Unity AI Gateway, MCP in Unity Catalog, Genie Ontology and Lakebase-backed memory. |
If you searched for "Databricks LakehouseIQ", the practical answer in 2026 is to look at Genie (and Genie Ontology) for business semantics and Agent Bricks for building agents.
The 2026 Databricks agent stack
| Need | Databricks component |
|---|---|
| Build and evaluate agents | Agent Bricks, with prompt optimisation, fine-tuning and supported agent harnesses |
| Ask questions of data in plain English | AI/BI Genie and Genie Ontology |
| Parse documents into tables | Document Intelligence SQL functions |
| Govern data, tools, models and agents | Unity Catalog and its registry |
| Control model and tool access and spend | Unity AI Gateway |
| Connect to external systems | MCP support in Unity Catalog (for example Google Drive, Jira, Slack, GitHub) |
| Agent state and memory | Agent memory service on Lakebase |
| Observe and audit | Agent traces and monitoring |
AI agents that help run the lakehouse
The original idea behind this post still holds: agents can take on much of the routine work of running a data platform. Typical patterns:
- Data pipeline agent. Watches job runs, explains failures from logs, and proposes or applies fixes such as re-runs or schema updates. Pairs well with agentic log analytics.
- Query optimisation agent. Finds slow or expensive queries and suggests clustering, caching or rewritten SQL.
- Governance agent. Checks new tables for missing owners, tags and sensitive columns, and drafts access policies for review.
- Insights agent. Answers business questions through Genie spaces and explains how each number was computed.
- Cost and capacity agent. Flags idle clusters and runaway workloads and recommends right-sizing.
How to build a governed agent on Databricks
- Pick one decision. For example, "approve or escalate a data access request". Define success and failure in measurable terms.
- Prepare context. Curate the tables, documents and metrics the agent may use, registered and tagged in Unity Catalog.
- Define tools. Expose functions, SQL and MCP connections with the narrowest permissions possible.
- Build and evaluate. Use Agent Bricks or your preferred harness; test on real past cases before any user sees it.
- Control. Route model and tool calls through the AI gateway with budgets and policies.
- Observe. Review traces weekly; feed confirmed errors back into evaluation sets.
For the broader tooling landscape, see our guide to MLOps, LLMOps and AgentOps platform tools, and for training and serving at scale, distributed machine learning.
Where Databricks agents fit, and where they do not
- Strong fit: agents whose main job is reasoning over governed lakehouse data, analytics questions, document extraction into tables and data platform operations.
- Needs more: agents that act across many business systems, run on factory floors or devices, or need to execute business processes end to end. These usually need an orchestration layer that spans the lakehouse, applications and edge. See agentic process automation.
Where it's heading: 2026 → 2030
| Period | Agents and the lakehouse |
|---|---|
| 2020 | Lakehouse for BI and ML; humans run pipelines and governance. |
| 2026 | Data platforms add agent builders, gateways, semantic layers and agent registries; agents assist with operations and analytics. |
| 2028 (expected) | Gartner expects 33% of enterprise software apps to include agentic AI; data platforms act as a control plane for many of them. |
| 2030 (our view) | Agents span data, applications and physical operations; cross-platform decision governance becomes a board-level requirement. |
How NexaStack helps
NexaStack complements Databricks by orchestrating agents and models across clouds, data centers and edge sites, using lakehouse data where it lives. Teams can run open models privately, connect agents to Databricks and other systems, and keep one view of traces, policies and cost. ElixirData, the decision harness for enterprise AI agents, adds decision records and policy checks when agents act on business outcomes.
FAQ
What is Databricks LakehouseIQ?
LakehouseIQ was announced in 2023 as a knowledge engine that learns an organisation's data and terminology. In 2026 those capabilities show up in Genie, Genie Ontology and Databricks Assistant rather than as a separate product.
What is Agent Bricks in Databricks?
Agent Bricks is Databricks' platform for building, evaluating and deploying AI agents on governed data, with model choice, support for popular agent harnesses, and built-in context and control features.
Can you build AI agents in Databricks?
Yes. You can build agents with Agent Bricks or open-source frameworks, govern them in Unity Catalog, and serve them through Databricks model serving and Apps.
Does Databricks support MCP?
Yes. Databricks announced MCP support in Unity Catalog in 2026 so agents can connect to external sources under the same governance.
What are the latest Databricks AI agent trends in 2026?
Agent platforms with gateways and budgets, semantic ontologies for business context, Postgres-backed agent memory, document intelligence in SQL, and tracing tied to governance.
Running agents across Databricks, your apps and your edge sites? Book a NexaStack demo to see one control plane for all of them.
