Databricks brought a long list of announcements to the Data and AI Summit, and they touched nearly every product on the platform. The premise behind it is one I buy. Agents read, loop, and write differently from the software we built for people, and the underlying data layer has to change to keep up. Databricks organized the whole event around context, control, choice, and cost.

Lakehouse RT takes a tier out of the stack

Lakehouse RT serves millisecond queries directly on the lakehouse, powered by a new compute engine Databricks calls Reyden. It matters because of what customers can stop running. Real-time serving has usually meant a separate, specialized store next to the lake, with replication keeping the two in sync, which adds cost, latency, and one more system to manage. Reyden runs inside Unity Catalog with the same governance as everything else, no separate copies and no change-data pipelines, so that whole tier comes out of the stack. It also gives Databricks the real-time speed that agents need in the moment, and both of the new applications they announced are built on it. That makes Lakehouse RT the foundation on which a lot of the agentic story now rests.

LTAP puts the transactional side on open formats

LTAP headlined the week for me, and it raised the most questions. LTAP pairs Lakebase, their serverless Postgres database for transactions, with the Reyden-powered lakehouse for analytics, and keeps both on a single copy of data in open formats so neither workload starves the other. The ambition isn’t new. People have chased unified transactions and analytics for well over a decade. What’s new is the transactional writes landing in open formats too, so operational data isn’t sitting in a proprietary store while only the analytics half stays open. For a buyer who doesn’t want to lock operational data to one vendor, that’s a genuinely different bet. The piece I’d still want their engineers to walk me through is what happens the instant a row commits and how both engines truly share a single copy without a sync step tucked in the middle. Get that right, and it’s a real step forward.

Genie Ontology matters more than the new Genies

Genie grew from a way to query your data into a full family of tools, with the business-user tier offered at zero per seat. No surprise there. Underneath it sits Genie Ontology, the context layer the whole family runs on, and that’s the piece I find most interesting. Two things I’ve long wanted to see more of from Databricks are richer active metadata and stronger loops for validating what gets fed to an AI, and Genie Ontology is the most direct answer they‘ve put forward on both. It learns from how people use the data and blends it with a curated layer, so an agent has a better shot at understanding what the business means by a word like “margin”. Where I’d still like to see more is trust. Ranking context by authority leans on the answer people use most, and I’d want to see how it catches one that’s confidently wrong before trusting it with the biggest calls.

Unity AI Gateway is really about cost

The Unity AI Gateway governs models, agents, and tools from one place, including agents running on other vendors’ platforms and third-party coding tools. For me, the cost controls are the most useful part. You can set budgets per employee or per agent and have the system fall back to cheaper models on its own. The mood in a lot of enterprises has shifted from moving too slowly on AI to running up a bill nobody can explain, and that’s where plenty of agent projects are stalling. Putting those controls in the governance layer, where the spending happens, is the right home for them. Omnigent, an open-source meta-harness that sits over other coding-agent harnesses, brings the same governed approach to how developers run their agents.

CustomerLake moves Databricks up the stack into the CDP

CustomerLake might be the most underrated announcement of the week. It puts Databricks in the customer data platform business, with profile agents that turn raw data into Customer 360 records and campaign agents that build audiences and run activation on top of them. That’s a real step up the stack. For years, Databricks sold the data foundation and left the customer-facing work to the application vendors. The argument now is that customer data already wants to live in the lakehouse for governance and cost reasons, so the profiles and the agents acting on them belong there, too. If it lands, it pulls a high-value workload away from the platforms that have owned it, and it puts Databricks in direct competition with companies it has partnered with. That’s a more aggressive expansion than most of what else they announced, and I’m curious how customers and partners take it.

Lakewatch and a serious push into security

Security is where Databricks is moving with the most intent right now. Lakewatch is their open take on SIEM, built to pull security data into the lakehouse rather than a closed system priced so high that teams toss data they should keep. At the Summit, they put real money behind it by acquiring Panther, an AI SOC platform with a deep library of integrations and detection-as-code, in addition to their earlier Antimatter and SiftD deals. That’s three security acquisitions behind one idea. Leadership has been open about wanting Databricks to become more of a security company, and the moves match the words. The pitch is a security data lake with agents handling the triage and investigation that a SOC can’t keep up with by hand. With Lakewatch in the SIEM and Panther in the SOC, Databricks now has both halves of a data security business, and my read is that they‘re just getting started. Even with established players in the category, I’d watch closely for what they build or buy next here.

Databricks is extending its lead with real work ahead

The bigger play here is familiar to anyone who follows Databricks. They want to be the governance, context, and cost layer for a customer’s whole data estate, not only the parts that run on their own platform, and for most of the week, they served that goal. Governing other vendors’ agents, reaching data wherever it lives, and giving the business-user interface away for free all serve it. On that ground, they‘re extending a lead they already hold. Security and marketing are the harder rooms because Databricks walks in as the newcomer, and being the strongest data platform in the building doesn’t yet make it the strongest in security or marketing. I also don’t think those are the last rooms they‘ll walk into. Databricks keeps pushing into the next adjacent market, and there’s no shot they‘re done, even if I can’t tell you which one is next. Cost was the theme that ran through the whole week, showing up in the engine that removes a tier, the gateway that caps spend, and the pipelines LTAP takes out. Most of the market is still selling how to do more with AI, while Databricks spent a real share of its time on how to spend less, and for where budgets sit right now, that makes sense to me.

Plenty of this is early, and a good chunk is still in preview, which is normal for a summit this size, so the real test is the year ahead as customers put it to work. I’d love to see a real correction loop in the context layer, something that notices when an agent reads the business wrong and sets it right, plus a clear technical walkthrough of how LTAP holds to a single copy of data and some proof that the cost controls hold up on real production budgets. None of that takes away from the week. This event was one of the more complete and coherent showings I’ve seen from Databricks, a clear point of view with the product to back it, and on recent form, I’d bet on them delivering most of it.