13, July 2026
by: Grace McCarthy
Blog by Ramen Bhar, Principal Data Engineer, Arreoblue.
Part 1 covered the full list of announcements, Part 2 the five that matter most. This last one is the most opinionated of the three: my own reading of the pattern, based on what was actually shown and shipped.
Put the last three summits side by side, and a pattern emerges.
What that table doesn’t quite capture is how connected these three years actually are. Databricks has spent them working through a single plan: stop being a data platform, and become the whole system instead, absorbing the operational database, real-time serving, security, marketing tech, and app hosting along the way, all in service of getting AI agents to work reliably in production.
The prize has shifted from the data itself to the context around it. Getting the data sorted was the job in 2024. Completing the operational stack occupied 2025. This year, all of that groundwork is being directed at agents, and what actually determines whether an agent is any good has turned out to be neither the model nor the raw data. It’s the layer that tells the agent what the data means and which sources to trust, which is exactly what Genie Ontology and Unity Catalog are quietly doing underneath everything more visible.
Databricks keeps bringing the system next door in-house. Table formats got folded in during 2024, and the operational database in 2025. This year, it’s happened to real-time serving, the sync pipeline itself, security, marketing tech, and app hosting, more or less simultaneously. Running fewer separate systems is cheaper in the long run, even if it means giving up some best-of-breed choices along the way.
Governance has had to keep up with what agents actually do. Unity Catalog began by controlling who could see which data. That got more granular in 2025. Now Unity AI Gateway is dealing with something harder: what an agent can do on its own, what it’s allowed to spend, and when a human needs to sign off. Deciding who can view a table is a much simpler problem than that.
Everyone is worried about cost, and Databricks knows it. This was the year the company said outright what had mostly been implied before: running agents at scale is going to get expensive. Most of what’s on this list responds to that in some way, whether it’s removing duplicate systems, capping and routing spend automatically, keeping an eye on what agents are doing, or offering cheaper custom models as an alternative to the default.
Open it up, then lead it. This one runs through all three years. Delta Lake and MLflow came first, then Unity Catalog went open source in 2024, and this year OpenSharing and Omnigent are doing the same job. The approach doesn’t really change: release something for free, let the market settle on it as the standard, then sell the managed, governed version to whoever eventually wants the support. It’s worked for them before, and there’s no obvious reason this round would be any different
Mostly, Databricks now has a hand in a much wider set of decisions than it used to, which makes this a reasonable moment to check your own roadmap against it. Plenty of teams still think of Databricks as somewhere data lives and pipelines run, which, increasingly, is the least interesting part of what it does. The harder question is whether your Ontology, your metrics, and your glossary are good enough to hand an agent and trust what comes back. Most companies aren’t there yet, and that’s just where the industry currently sits.
Worth being honest, too, about how early some of this is. LTAP, CustomerLake, Lakewatch, and much of Omnigent are still in private preview, beta, or “coming soon.” None of that’s a reason to ignore them, but it is a reason to treat the roadmap as a roadmap rather than a shopping list, and build your plans around what’s actually shipped.
That’s the series. If you have questions about any of this for your own setup, you know where to find us.
If you are interested in finding out more, please contact us.
Have a question or want to make an enquiry?
Contact us
