in Blog
July 16, 2026
Databricks Data + AI Summit 2026: New Features Explained
Databricks Data + AI Summit 2026 marked the shift from building AI agents to operating them at enterprise scale. The headline moves: Genie One went generally available as a cross-department AI coworker, LTAP proposes killing the ETL pipeline entirely, Unity AI Gateway adds hard spend caps to stop runaway agent costs, and Databricks made its third security acquisition (Panther) to back its new Lakewatch SIEM.
Below, every major announcement, what it is, why it actually matters, and where it stands today.
KEY TAKEAWAYS
Genie One, Databricks’ AI coworker, is now generally available across web, iOS, and Android.
LTAP proposes eliminating ETL/CDC pipelines entirely by unifying transactional and analytical data on one copy — but it’s announced as “coming soon,” not shipping yet.
Unity AI Gateway adds hard spend caps so an agent can be auto-stopped before it runs up an unexpected bill.
Databricks acquired Panther — its third security acquisition — to expand its new Lakewatch SIEM with 100+ pre-built integrations.
Several major features (CustomerLake, Genie App Builder, Genie ZeroOps) are still in private preview — worth tracking, not yet ready for production dependency.
What are Databricks’ new agent infrastructure tools?
100K+
Agents built on Agent Bricks
1Q+
Tokens processed per year
30K+
In-person Summit attendees
| Feature | What It Is | Why It Matters | Proof Point / Status |
|---|---|---|---|
| Agent Bricks | A developer platform for building, deploying, and governing production agents across any model or framework. | Solves the part of agent-building nobody talks about: 99% of the real work is deployment, security, evaluation, and monitoring — not the model call itself. | Customers include AstraZeneca, 7-Eleven, Fox Corporation, Block. Expanded platform announced June 16, 2026. |
| Agent Memory Service | Managed agent memory built on Lakebase, with short-term (session) and long-term (cross-session) tiers. | Previously, giving an agent memory meant standing up your own Postgres instance and checkpoint logic by hand. This collapses that into one API call. | Integrates natively with LangGraph checkpointers and the OpenAI Agents SDK. Preview. |
| Databricks Sandbox | Isolated virtual machines with downscoped Unity Catalog access for running agent code. | Lets an agent execute untrusted or generated code without touching production data or credentials. | Preview. |
| Omnigent | Open-source (Apache 2.0) meta-harness for running and combining multiple coding agents. | Prevents lock-in to one agent framework, swap or combine LangGraph, CrewAI, or Claude Code with a one-line change. | Open source, announced at DAIS 2026. |
How is Databricks eliminating the ETL pipeline?
LTAP pairs Lakebase (transactions) with the lakehouse (analytics) on one copy of open-format data — removing the sync step between operational databases and analytical tables entirely.
| Feature | What It Is | Why It Matters | Proof Point / Status |
|---|---|---|---|
| Lakebase | Serverless Postgres (OLTP), natively integrated with the lakehouse. | Gives agents a transactional, low-latency read/write layer for operational state — without leaving the governed lakehouse. | Generally available earlier in 2026. |
| Lakebase Search | Hybrid vector + full-text retrieval built directly into Postgres. | Removes the need for a separate vector database bolted onto an agent stack. | Beta |
| Lakehouse//RT | Real-time query engine (powered by Reyden) delivering sub-100ms latency directly on Delta/Iceberg tables. | Removes the separate real-time serving tier (e.g., Druid, ClickHouse) most real-time dashboards previously required. | Beta. PointClickCare reported ~30% faster average performance, 10x faster queries, vs. their prior warehouse. |
| LTAP | An architecture unifying OLTP and OLAP workloads on a single copy of open-format data. | The actual pitch to kill ETL/CDC pipelines — both workloads read and write the same open-format copy. | Announced as “coming soon” — architecture, not yet shipping. |
| Lakeflow (Connect, Jobs, Designer) | Unified data engineering stack: managed ingestion, orchestration with External Orchestration, and a no-code visual builder. | Consolidates pipeline tooling that used to require stitching together separate ingestion, orchestration, and transformation tools. | Lakeflow Designer added to Databricks Free Edition (500k+ users). |
| Spark Declarative Pipelines | Declarative SQL APIs for streaming, with a new Real-Time Mode. | Makes streaming pipelines as easy to write as a SQL query, with near-live latency. | Preview. |
| Zerobus Ingest | High-volume event ingestion with Kafka-compatible APIs. | Existing Kafka producers plug in without code changes, while Databricks manages the cluster. | Generally available.. |
What’s new in the Genie family of AI agents?
| Feature | What It Is | Why It Matters | Proof Point / Status |
|---|---|---|---|
| Genie One | An agentic AI coworker for business teams across web, iOS, and Android. | Non-technical users can direct multi-step work, grounded in governed company data instead of guessing from documents. | Generally available. Grounded in Genie Ontology, which learns from 50+ connected apps. |
| Genie Ontology | A live, self-improving business-context layer that learns definitions (e.g., what “ARR” means at your company). | Stops an agent from hallucinating a metric definition — grounds every answer in the business’s actual meaning of a term. | Underpins Genie One and Agent Bricks. |
| Genie Code | An autonomous coding agent across Lakeflow and the ML lifecycle. | Lets a data engineer hand off a multi-step task and review a pull request instead of writing it by hand. | Generally available. |
| Genie Agents | Turns any Genie conversation into a reusable, shareable agent. | Removes the “rebuild from scratch” step every time a useful workflow needs repeating. | Generally available. |
| Genie App Builder | A “vibe-coding” tool turning a plain-language description into a working internal app wired to governed data. | Skips the traditional low-code builder step — describe the app, get a working, governed one. | Private preview. |
| Genie ZeroOps | A background agent that detects pipeline/table/model failures, runs root-cause analysis, and proposes fixes. | Cuts the manual triage step out of maintenance — someone doesn’t have to notice the failure first. | Private preview. |
Genie pricing is pay-as-you-go, not seat-based
How is Databricks governing agents that can take real actions?
| Feature | What It Is | Why It Matters | Proof Point / Status |
|---|---|---|---|
| Unity AI Gateway | Central runtime governance for agents: cost budgets, model routing, and reasoning-trace logging. | Answers the “agent racked up thousands in cloud spend overnight” problem — auto-stops requests at a budget threshold. | Includes Contextual Service Policies (beta): human approval required before actions like a GitHub push or sending an email. |
| Unity Catalog Semantics | A Glossary, Domains, and Metrics layer giving agents and humans one shared definition of business terms. | Prevents two agents (or departments) from computing “revenue” two different ways. | Includes cross-cloud, cross-region addressability. |
| OpenSharing | Delta Sharing’s evolution into an open, Linux Foundation–governed protocol. | Lets organizations share agent skills, models, and prompts across company boundaries without lock-in to one protocol. | Now a Linux Foundation project. |
| Governance Hub | A central command center in Unity Catalog for tracking data risk and external API usage. | Gives a security officer one view instead of stitching together logs from a dozen systems. | Generally available. |
What did Databricks announce for ML and vertical-specific platforms?
| Feature | What It Is | Why It Matters | Proof Point / Status |
|---|---|---|---|
| AI Runtime | Serverless, pay-as-you-go NVIDIA GPUs (A10, H100) attachable directly in notebooks. | Removes cluster-configuration overhead — attach a multi-node GPU and pay only for active training time. | Generally available.. |
| CustomerLake | An agentic CDP built directly into the lakehouse, using Profile and Campaign Agents. | Replaces the “copy data into a separate martech tool” step — campaigns run on data that never leaves Unity Catalog governance. | Private preview. Early customers: HP, Circle K, AB InBev, Getnet by Santander. |
| Document Intelligence | SQL functions (ai_parse_document, ai_extract, ai_classify) for parsing documents natively. | Turns unstructured PDFs and manuals into queryable data without a separate document pipeline. | Generally available. |
| Lakewatch | Databricks’ agentic SIEM, storing security telemetry in open lakehouse format. | Legacy SIEMs’ per-GB pricing pushes teams to pre-filter logs, creating blind spots. Lakewatch removes that incentive. | Launched March 2026 at RSA Conference; expanded at DAIS 2026. Built on the Open Cybersecurity Schema Framework. |
| Panther | An AI SOC platform Databricks agreed to acquire, adding 100+ integrations to Lakewatch. | Gives Lakewatch immediate breadth instead of building integrations from scratch. | Announced June, 2026 — Databricks’ third security acquisition after Antimatter and SiftD. Anthropic is a named customer. |
FAQ
What is LTAP and is it available now?
LTAP (Lake Transactional/Analytical Processing) is Databricks’ architecture for running transactional and analytical workloads on one copy of open-format data. As of the summit, it was announced as “coming soon” — not yet generally available.
Is Databricks entering the cybersecurity market?
Yes. Lakewatch (its agentic SIEM, launched March 2026) is its own product, and the planned Panther acquisition is its third security acquisition, following Antimatter and SiftD.
What's the difference between Lakebase and LTAP?
Lakebase is the serverless Postgres product itself (the transactional engine). LTAP is the broader architecture that pairs Lakebase with the lakehouse so both workloads read and write the same data.
Category:
Data Engineering
