By Nicholas Anile | AI/ML Principal, Mantel
Key takeaways for business leaders:
- The platform race is settling; the new competitive frontier is the layer above it. Databricks framed this year’s summit around four core pillars: context, control, cost, and choice. The message is direct: models are already good enough. Winning organisations will be those that effectively manage how enterprise knowledge reaches agents, govern what those systems can do, and prevent AI spend from being prohibitively expensive at scale.
- Headline releases point towards a single, governed foundation for operational systems, analytical data, and AI agents. Omnigent unifies multi-vendor developer tools via an open-source meta-harness, while Agent Sandboxes allow autonomous code execution without risking live environments. Lakebase, Reyden, and LTAP eliminate the costly ETL pipelines that traditionally silo operational data, and Genie Ontology maps enterprise knowledge into a structured graph to cut token costs and search latency.
- Organisations managing complex, siloed data layers should pay close attention. While local Australian availability currently varies (Omnigent is in beta, Lakehouse//RT is a gated beta, and the Sandbox capability is not yet available), the trajectory is clear. Databricks is building towards a lakehouse the business actively runs on.
We were at Databricks Data + AI Summit 2026 in San Francisco, and the mood was clear: the platform race is settling. What matters now is the layer above: how enterprises manage context, cost, control, and choice. Here’s what caught our attention.
The theme across both keynote days was simple: models are already getting good enough, and the actual gap is managing context. Ali Ghodsi framed all announcements at the conference around four problems, the four C’s:
- Context: Getting your data and its meaning into the tools and agents that use it
- Control: Governance and identity across the platform
- Cost: AI spend, he warned, turns prohibitive for most companies within 6 to 12 months; and
- Choice: no lock-in to one model, cloud, or vendor.
The releases spanned the whole platform, from the data, analytics, and AI/ML foundation, through governance, to the agents on top, and they pointed the same way: the lakehouse is becoming the place where the business runs, not just where its data sits. Many of the announcements drive at making enterprise workloads easier on Databricks.
We loved being there to support some of our customers who presented this year, including NAB, Alinta, and Suncorp. This blog highlights just a few of the many announcements and 800+ sessions. Our team certainly got their step count up, shifting between all the venues of Moscone Centre and the wider conference area.
Omnigent: The latest harness
One of my personal highlights was the unveiling of Omnigent, a new open-source ‘meta-harness’ designed to establish a standard infrastructure framework for agentic coding tools. While technical teams have spent the last few years building custom workflows around their preferred tools, testing ideas across different harnesses or collaborating seamlessly has remained a major challenge. Many devs know the infuriating headache of kickstarting a local agent only to realise they have to leave their laptop open for it to finish running. It strongly reminds me of my student days running local scripts. Omnigent aims to change that.
Rather than acting as another full-stack agent framework, Omnigent operates as a unifying API layer above existing tools (like Claude Code, Codex, and Cursor). It directly addresses the three core challenges of enterprise agent management:
- Design: It allows developers to combine models and harnesses in parallel, such as running a cheaper model for bulk work and calling a premium model only when it gets stuck.
- Collaboration: It introduces live session sharing, allowing teammates to read, comment on, and direct agents in real time. Allowing developers to collaborate directly on the code will be a huge experience improvement.
- Governance: Omnigent offers a new style of security policies. Rather than binary controls, it uses a contextual approach. For example, an agent that has just accessed a confidential document can be blocked from sending outbound emails, while email tools stay enabled in lower-risk contexts. Also critical is the ability to set spend budgets for agent workloads.
Omnigent is yet another open-source initiative that Databricks is deeply committed to championing. By establishing a transparent, open standard, they are inviting the global community of developers, organisations, and research labs to collaborate, hopefully triggering a powerful network effect that will drive innovation far faster than a closed system could. Omnigent doesn’t require the Databricks platform to function; it is entirely infrastructure-agnostic, allowing organisations to deploy the runner and server nearly anywhere.
Agent Bricks and the sandbox: letting agents run code without the risk
Agent Bricks is Databricks’ platform for building agents, and it’s already running at scale with more than 100,000 custom agents built on it by Databricks customers. It gives you your pick of frontier model (now including Grok and Kimi) plus a suite of tooling to build agentic solutions. The piece we found pretty exciting in this capability space is the new Sandbox offering.
An agent is one of the most privileged entities you can put in your stack, and much of what agents do is write and run code. Running model-generated code straight against your external systems or deploying into your environments can be risky. A sandbox gives each task its own disposable environment that can start in seconds, with the leading capabilities of Unity Catalog to govern access to your data that stays inside your workspace boundary. They’re designed to be cheap and fast enough to spin up per task (especially useful for development contexts). We’ve grappled with exactly this: embracing how useful agents are, while wanting to govern them with some of our more sensitive customer environments. Let agents act, but contain them.
”An agent is one of the most privileged entities you can put in your stack, and much of what agents do is write and run code. Running model-generated code straight against your external systems or deploying into your environments can be risky. A sandbox gives each task its own disposable environment that can start in seconds, with the leading capabilities of Unity Catalog to govern access to your data that stays inside your workspace boundary.
Nicholas AnileAI/ML Principal, Mantel
Lakebase, Reyden and LTAP: operational and analytical data, without the copies
Three releases line up into one story, and it is a monumental shift in the database space that we should pay attention to.
Lakebase is the fully managed PostgreSQL database that sits on the lake (released at DAIS last year), creating the opportunity for transactional data and low-latency operational writes to fit within the Lakehouse. Reyden is a newly announced engine behind Lakehouse//RT, optimised for the high-concurrency, low-latency patterns that operational workloads and agents need. It enables millisecond queries against the tables without a separate serving layer or distinct data service. LTAP ties the two together: it uses spare capacity in Lakebase’s caching layer to convert what you write into a columnar form in the background. As a result, every table is available in the lakehouse with no Change Data Capture (CDC) pipelines to maintain, zero risk of data going stale, and all the benefits of longer-term lake storage.
The three work together: write to Lakebase, LTAP lands that data in the lakehouse automatically, and Reyden serves it in milliseconds. You get one governed copy, no ETL jobs, and no stale data discrepancies between your operational and analytical systems. Lakebase also brings a whole bunch of capabilities that are incredible for agents such as database cloning and forking – enabling rapid experimentation and teardowns without the fear of data loss, as snapshot rollbacks are instantly possible.
Elevating your context with Ontology
Another headline release was Genie Ontology, a new capability within Databricks designed to map enterprise knowledge into a structured knowledge graph. The feature runs on a new algorithm called OntoRank, which automatically categorises data and determines its relative importance during the graph-building process. This allows platform agents to surface highly relevant information much faster. The performance metrics shared were notable: agents grounded by Genie Ontology showed higher accuracy and required fewer runtime searches, reaching outcomes faster with lower token consumption. This context layer looks poised to become a vital new foundational pillar for the modern AI stack.
My approximation of how a graph can be useful for a discovery agent below:
The rollout for these features is actively moving forward, with varying timelines for our region:
- Omnigent: Currently in beta and fully available in Australia.
- Lakehouse//RT: Available in Australia via a gated beta.
- Sandbox: In beta, but not yet available in Australia.
- Genie Ontology: Public Preview.
The releases at this year’s summit point in a consistent direction.
Databricks is building toward a world where your operational systems, analytical layer, and AI agents share the same governed foundation. For organisations that have spent years managing the cost and complexity of keeping those layers separate, that’s worth taking seriously.
If you want to talk through what any of these changes mean for your environment, reach out to our team to start a conversation.
