Databricks at Data + AI Summit 2026: The Releases, and How to Read Them
What Databricks announced this year, what each release is for, and where Substrate works with it.
Data + AI Summit 2026 brought a major wave of releases, and they point in one direction: Databricks is building out the layer where enterprises build, govern, and run AI agents on top of their own governed data. The work is coherent and anchored to the platform's core strength: keeping data unified and governed in one place.
Here is a straightforward read of the releases that matter most, what each is for, and where Substrate plugs in.
Agent Bricks: a platform for building agents on governed data
A platform for assembling production agents: any model, any harness, an agent memory service on Lakebase, Document Intelligence, the Genie Ontology for grounding, and a secure sandbox. The grounding in governed enterprise data is the standout, the part most agent platforms leave to the customer. Databricks put the framing plainly: the core agent loop is about 1% of the work, and the surrounding infrastructure (memory, evaluation, context, governance) is the other 99%.
Outcomes
- Take a support agent from prototype to a production accuracy and escalation bar, wired into the existing ticketing system and proven against real cases.
- Stand up a document-processing pipeline that holds an extraction-accuracy target across millions of inconsistent, real-world contracts or claims.
- Roll an internal "ask the data" agent out org-wide, with the access controls, grounding, and monitoring needed to trust its answers.
Substrate registers Agent Bricks' tools and Document Intelligence as skills and builds on its Lakebase-backed memory, running them inside its own engineering loop.
Unity AI Gateway: runtime governance for models, agents, and tools
A governance plane that extends Unity Catalog past data to the runtime interactions of AI. Register and govern hosted and external models, agents, and skills in one catalog; set policies that allow, deny, or require approval for sensitive actions; cap spend; route across models; and land end-to-end traces in the lakehouse. The distinctive part is governing agents and skills in the same catalog as the data they touch.
Outcomes
- Design the org-wide policy model (what each agent may and may not do, mapped to real risk) and operationalize it across teams.
- Bring shadow model and agent usage under governance without disrupting the teams already relying on it.
- Stand up a cost-governance program (caps, routing rules, and chargeback) that actually holds across the organization.
Substrate runs inside the Gateway. Its routing, spend caps, and approval policies apply to the engine directly, and its traces land back in the lakehouse alongside everything else.
Unity Catalog: Glossary, Domains, Metrics, and the Genie Ontology
The catalog gains a real semantic layer: a Glossary of authoritative terms, Domains that scope business context for agents, and governed Metrics for consistent KPIs, plus the self-learning Genie Ontology that teaches the platform the organization's own meaning, and a four-level cross-cloud namespace so one asset has one address everywhere. This is what lets AI reason in the business's language rather than guess at it.
Outcomes
- Reconcile conflicting metric and term definitions across departments that disagree today, then encode them so every dashboard and agent finally agrees.
- Migrate a legacy catalog and its governance into Unity Catalog with lineage and access intact.
- Build the domain and access model that scopes each agent's context correctly across business units.
Substrate reads the catalog (Glossary, Domains, Metrics, and the Genie Ontology) as ground truth. The better governed the catalog, the more accurate everything built on it.
Omnigent: an open meta-harness for agents
Open-sourced under Apache 2.0, Omnigent sits above individual agent harnesses and wraps them in a uniform API, with isolated sandboxes, access controls, cost ceilings, and human-in-the-loop approval. A managed version runs on Databricks. It is open and governance-first, a standard the ecosystem can build on rather than a single-vendor silo.
Outcomes
- Consolidate fragmented agent tooling onto one governed harness layer, migrating existing agents across.
- Build the sandbox and approval framework that lets agents run unattended within boundaries the org trusts.
- Operate an unattended automation loop on top of Omnigent that ships real work, not just demos.
Substrate adopts Omnigent as its harness layer and calls agents through its uniform API, so the harnesses a team standardizes on are the same ones the engine runs.
CustomerLake: an agentic customer data platform
A CDP built into the lakehouse: Customer 360, identity resolution, audience building, and activation, with agents reacting to customer context in real time. It brings marketing and CX data into the same governed place as everything else.
Outcomes
- Migrate and reconcile customer data from fragmented legacy systems so identity resolution works on real, dirty data.
- Build the activation integrations to the specific downstream marketing and sales tools the business actually runs on.
- Stand up and tune the real-time campaign logic around the company's actual customer journeys.
Where an organization adopts CustomerLake, Substrate builds the identity logic, the audience pipelines, and the activation integrations around it.
Lakebase and the data foundations
The operational core beneath the agent layer: Lakebase, the transactional layer agents persist to; LTAP, which unifies transactional and analytical data on one copy; and real-time serving for sub-second workloads.
Outcomes
- Migrate operational workloads onto Lakebase and LTAP without breaking the applications that depend on them.
- Re-architect a pipeline to hit a sub-second serving target for a live, customer-facing product.
- Build the production-grade state and memory layer an agent fleet relies on to run reliably.
Substrate persists and verifies state on Lakebase, and uses its copy-on-write branching to test against a full-fidelity branch of live data.
Also notable
- OpenSharing: Delta Sharing evolved into a Linux Foundation standard that now covers models, skills, and unstructured data alongside tables: more portability, less lock-in across a mixed estate. Substrate consumes shared assets as governed context without copying them.
- Next-generation Genie: cross-source natural-language analytics, now on web and native mobile.
- Spatial SQL, GA: native geospatial types and 90+ ST_* functions for the teams that need them.
- Lakehouse//RT: sub-second serving for real-time dashboards.
- Grok on Agent Bricks: xAI's models join the native model choices.
How to read it all
Two threads run through the release set. Databricks is enclosing the full agent lifecycle: build with Agent Bricks, govern with Unity AI Gateway, ground with Unity Catalog and the Genie Ontology, all anchored to governed data. And openness is deliberate: Omnigent and OpenSharing both favor portability and shared standards. For any enterprise on the lakehouse, the foundation is stronger and more coherent than it was a year ago.
Substrate is the engineering layer that runs on top of it: it adopts Omnigent as its harness, reads Unity Catalog as ground truth, runs within Unity AI Gateway, persists to Lakebase, and registers Agent Bricks' tools as skills. The platform supplies the governed surfaces; the engine runs the work against them.
Sources: Databricks blog and newsroom, June 2026; the Linux Foundation; xAI; and third-party summit coverage including TechTimes, MarTech, Pulse 2.0, and Qubika. Release availability is as stated at the summit and subject to change.
Ready to transform
your enterprise?
Let's build something that lasts. Our team is ready to talk.