A clear, product by product walk through of Data and AI Summit 2026, and why context is now the center of the platform.
This year, 31,309 people filled the Data and AI Summit. It is now the largest data and AI conference in the world. The room was big, but the bigger story was the shape of the announcements.
For years, a Databricks keynote was mostly about data. This year, almost every product was about agents, and about one quiet idea underneath all of them. The hard part of enterprise AI is no longer intelligence. The hard part is context. Getting your real business meaning to the machine, in a way it can trust and act on.
Before we walk through each product, a small note of credit. Almost everything Databricks ships still grows from open source. Spark, Delta Lake, MLflow, Unity Catalog, and now an open embrace of Postgres and Iceberg. They keep the foundation open while moving fast. That is rare, and it matters.
[!notice] This summit was not really about new features. It was about one idea. The data platform is becoming the place where agents go to understand your company. Read the announcements through that lens and they all line up.
Let us go product by product.
Genie used to be a chat box over your Databricks data. This year it became a family.
Genie One is the headline. It is an agentic coworker for business teams in finance, sales, marketing, and operations. It works across any data, structured or unstructured, analytical or operational, inside Databricks or in connected apps. It does not just answer questions. It can produce reports, charts, alerts, and documents, and it is available on web, iOS, and Android.
The interesting part is why it works better than older AI helpers. Genie treats governed enterprise data as the ground truth, not loose documents or embeddings. So when a CFO asks why margins changed, Genie looks up the real answer in trusted tables through SQL, instead of guessing from scattered files.
Genie Ontology is the engine behind that. It is a live context layer that continuously learns your business from data, docs, tags, tickets, chats, meetings, and more than 50 connected apps like Google Drive, Slack, Jira, and SharePoint. It captures what a churned customer means, when your fiscal year starts, and which tables matter most. This is the context layer the whole summit kept pointing back to.
Genie Agents let any good Genie conversation be saved as a reusable agent. It keeps the sources, instructions, and behavior, so a coworker can call it by name and repeat a trusted workflow.
Genie App Builder is a managed vibe coding environment for the enterprise. You describe what you need, and it generates a working app connected to real governed data, with Unity Catalog permissions built in from the start.
Genie Code is an autonomous agent for data teams. It helps plan, build, and run data engineering, machine learning, and analytics work, with a workspace to track progress and switch between projects.
Genie ZeroOps is the one that data engineers will feel most. It is a background agent that monitors your pipelines, jobs, tables, and models, investigates when something breaks, and proposes a fix. Think of the broken 2 AM pipeline that now gets investigated before you wake up.
Genie has no seat-based pricing, and organizations get up to ten dollars free per user each month. Genie One, Genie Agents, and Genie Code are generally available. Genie App Builder and Genie ZeroOps are entering private preview after the summit.
If Genie is for business teams, Agent Bricks is for developers. Last year it launched as a way to build quality agents. This year it became a full agent platform, because Databricks learned something honest.
The core agent loop is only one percent of the work. The other ninety nine percent is the hidden plumbing. Token capacity, deployment, security, evaluation, monitoring, context, and sharing. Developers were stuck building infrastructure instead of agents.
Agent Bricks now solves that across three needs.
Choice. Use any model and any harness. Frontier models from OpenAI, Anthropic, and Gemini, open models like Qwen and Kimi, and now Grok through a new SpaceX partnership. You can also train custom models on your own data. Databricks showed a custom data agent that matches frontier models on Genie tasks at much lower cost per query. For harnesses, it supports LangGraph, CrewAI, Agno, the Claude Code SDK, and the OpenAI Agent SDK, plus a managed version of Omnigent, the open source meta harness they released over the prior weekend to orchestrate harnesses together.
Context. This is where the data platform earns its place. Agent Bricks connects agents to data everywhere through MCP, grounds them in Genie Ontology, and ships built-in Databricks agent tools, including a document search that is now three times faster. It adds an agent memory service powered by Lakebase, Document Intelligence functions in SQL for parsing PDFs and images, and Databricks Sandbox for running agent code in safe, isolated VMs.
Control. Agents are some of the most privileged actors in a company, so they need governance. That brings us to Unity Catalog.
This was the deepest set of updates, and the most important for our world. The catalog has moved from a system of record to a runtime decision-maker. Databricks framed it as three jobs: control, context, and choice.
Control. The new Unity AI Gateway governs models, agents, tools, and MCP servers in one place, the same way you already govern tables. It adds Contextual Service Policies so you can allow, deny, or require approval for specific actions, like writing to a sensitive folder. It adds spend caps across providers and unified agent tracing. There is also a Governance Hub command center, and stronger access control through ABAC grant policies and new role based access control, so you can model things like exclusive access for clinical or country-specific data.
Context. Unity Catalog Semantics gives agents and people one shared source of meaning. New pieces include Glossary for authoritative business terms, Domains (now public preview) to organize assets into business areas so an agent sees a scoped, relevant slice instead of the whole catalog, and Metrics for defining KPIs like revenue and churn once and querying them everywhere. Metrics is open source and works in Apache Spark. This user-defined layer is what feeds Genie Ontology. External lineage is now generally available and stretches the lineage graph beyond Databricks, and Table Insights now shows column-level popularity.
Choice. Unity Catalog now reaches across clouds, regions, and accounts with a new four-level namespace, so one asset has one address everywhere. There is managed disaster recovery for failing over in minutes, broader cross-format support including Apache Iceberg v3 (now GA) and a new FILE type for governing PDFs, images, audio, and video, and OpenSharing, the next step beyond Delta Sharing, now a Linux Foundation project for sharing data and AI assets across vendors.
Databricks Lakeflow is now generally available, bringing unified ingestion, transformation, and orchestration under Unity Catalog. This includes Lakeflow Connect for automated ingestion from over 100 sources and Lakeflow Declarative Pipelines (formerly Delta Live Tables) for building reliable data flows.
Lakebase, the managed Postgres foundation, is now generally available. It powers the new agent memory service in Agent Bricks, so agents can keep session history and context across runs. This is the operational, real-time side of the platform sitting next to the analytical side. We wrote a full piece on why that matters in the Lakebase article.
Databricks also announced Lakehouse//RT, which brings real-time analytics directly to the lakehouse powered by a new compute engine called Reyden. It delivers millisecond query latency directly on governed Delta Lake and Apache Iceberg tables. The theme here is the same as Lakebase. The line between live operations and trusted analytics keeps getting thinner, on one platform.
With CustomerLake, Databricks entered the marketing industry with an agentic customer data platform. It is built on the same governed lakehouse, so customer data, analytics, and agents share one foundation instead of yet another disconnected tool.
Databricks agreed to acquire Panther, building out the security lakehouse category, and tied it to Lakewatch, its lakehouse-native security platform. As agents become privileged actors, their traces and actions need to be watched and audited like everything else. Security stops being a separate silo and moves into the lakehouse.
Finally, third party apps are now available on Databricks Marketplace, alongside Genie Sharing for collaborating on agents across organizations. The platform is starting to look like a place where you buy, share, and transact, not only store and query.
Read all of this together and a pattern appears. When agents can write a pipeline, fix it, and answer questions over it, the durable skill is not memorizing one more tool. It is understanding the systems underneath.
Governance. Data modeling. Freshness. Lineage. How operational and analytical data relate. The exact things that make an agent trustworthy are the exact things a strong data engineer already cares about.
The summit, in plain terms, said that context is the bottleneck. And context is built by people who understand data deeply. That is good news for anyone willing to learn the fundamentals well.
If this future excites you, the way in is not to chase every product name. It is to build the base that makes all of them make sense.
Start with the first lessons to see how we think about data engineering. Spend real time with Unity Catalog and governance, because governance is now the center of the agent story. Learn the medallion architecture so you understand how trusted data is shaped. And when you are ready to prove it, walk through the certification guide.
Keep reading next: the Lakebase deep dive and the AI flywheel to see where the platform is heading.
The future will belong to engineers who can give machines trustworthy context. That starts with understanding your own data, deeply and calmly, one system at a time.