About the author: Ido Bronstein is the co-founder and CEO of Upriver, the AI data engineering platform that connects to a team's full data environment - warehouse, orchestrator, and code - and executes data engineering work end-to-end with results you can actually trust. Connect on Linkedin.
The short version
- Strip the logos off the slides and Databricks and Snowflake made the same bet in 2026: not to build the best AI agent, but to be the trusted data foundation every agent runs on.
- The lake-format war is over - both shipped Apache Iceberg V3. Competition moved one layer up, to the context (semantic) layer.
- Genie Ontology vs Horizon Context is the real fight: whoever owns the semantic layer owns the agent's trust.
- The interface is going agentic (CoCo/CoWork vs the Genie suite) - and routing agents through it earns each vendor a cut of every token.
- Databricks took a real swing at the 40-year OLTP/OLAP divide with Lakehouse//RT. Snowflake has no counter yet.
- Bottom line: your platform choice is now a foundation choice - you're picking whose context engine, interface, and governance your agents will live on.
The 2026 summit season gave the clearest signal yet of where the data industry is going. Within two weeks, the two largest data platforms each laid out a roadmap built around the same idea: agentic data engineering. Using AI agents to do the work of preparing, governing, and serving data, with the platform acting as the trusted foundation those agents run on.
This report gathers what the major players announced, describes where each one stands today, and looks at where the broader market is heading. The more useful question for data teams is not who "won" a keynote, but what the shared direction means for how data work gets done over the next few years.
What "agentic data engineering" actually means
Agentic data engineering is the practice of putting AI agents in the loop for the tasks that data engineers have always owned: modeling data, writing and maintaining pipelines, defining metrics, monitoring quality, and answering questions against governed data. Instead of a person writing every transformation by hand, an agent proposes and executes the work, while the engineer sets direction, reviews output, and stays accountable for what ships.
For that to work, three things have to be in place: a context layer that tells the agent what the data means, an interface where humans and agents collaborate, and a governance model that keeps agent behavior, identity, and cost under control. Almost everything announced at the 2026 summits maps onto one of those three needs.
Where Databricks stands, and where it's heading
At its Data + AI Summit (June 15–18, 2026), Databricks positioned itself as the technical foundation for teams that build their own agents. Its tagline for the event - the database AI agents deserve - captured the emphasis on depth and engineering control.
What they announced. The most discussed release was Genie Ontology, a self-improving context engine that learns what an organization's data means as people use it, so agents can reason about the business rather than just the schema. On the development side, Agent Bricks matured into a full platform for building agents on any model with any framework, running in isolated sandboxes; Databricks reported that more than 100,000 agents have been built on it, processing over a quadrillion tokens a year. Governance arrived through the Unity AI Gateway, which applies cost controls, routing, and spend limits across models, agents, and connected services, and which Databricks built to integrate outside security vendors rather than replace them. The company also released Omnigent, an open-source framework for running agents across harnesses, alongside a managed version.
Databricks also pushed on infrastructure. Lakehouse//RT, running on a new vectorized engine, targets sub-second analytics directly on the lake, and Lakebase brings serverless Postgres with instant database branching into the platform. Apache Iceberg V3 reached general availability.
Where it's heading. Databricks is moving toward a single substrate where transactional and analytical workloads, agent development, and governance all live together. Its consistent theme is openness - open table formats, open agent frameworks, open integration - paired with managed services that monetize the operation of that openness. The direction of travel is a platform that handles the infrastructure of agentic data engineering so that engineering teams can focus on what their agents should know and do.
Where Snowflake stands, and where it's heading
At Snowflake Summit (June 1–4, 2026), under the theme "Making AI Real for Business," Snowflake emphasized breadth and ease of adoption. Its message - bring agentic AI to all your data - pointed at reducing the effort required to put AI to work across an organization's existing estate.
What they announced. Snowflake leaned into a context-and-interface strategy. Horizon Context, paired with Semantic Studio and Semantic View Autopilot, lets teams define shared business logic without deep SQL expertise and keeps those definitions current. The company renamed and expanded its agent products: CoWork (formerly Snowflake Intelligence) and CoCo (formerly Cortex Code), the latter now a native desktop application that can execute data tasks on its own using a catalog of reusable skills. Cortex Training lets customers customize open-weight language models on fully managed GPUs, and Cortex Sense supplies agents with context through managed connectors.
On governance, Snowflake took a more self-contained approach: Agent Identity gives each agent a verified identity, the Trust Center adds observability, and machine-learning detection guards against prompt injection. Interoperability came through the Horizon Catalog, built on Apache Polaris, with general availability of Apache Iceberg V3 and bi-directional read and write. The company also introduced Snowflake Datastream, a managed streaming service, and reinforced its commercial position with a multi-year, multi-billion-dollar cloud commitment and an acquisition aimed at agent identity.
Where it's heading. Snowflake's trajectory favors simplicity and reach: make it straightforward to bring AI to data wherever it sits, with managed services that hide complexity and a governance layer kept largely inside its own walls. Its bet is that most organizations want agentic data engineering to be turnkey rather than something they assemble themselves.
The shared direction across the market
Looking past any single product, four shifts are now common across the major players, and they define the near-term state of agentic data engineering.
The first is that open lake formats have become a settled standard. With both platforms shipping Apache Iceberg V3, the table format is no longer a point of differentiation. Competition has moved up the stack.
The second is that the context layer is now the center of gravity. Genie Ontology and Horizon Context are different expressions of the same conviction: an agent is only as trustworthy as its understanding of what the data means. The semantic layer, long treated as optional tooling, has become foundational infrastructure.
The third is that the interface to data is becoming agentic. Both companies are investing in agent-driven surfaces - CoCo and CoWork on one side, the Genie suite on the other - reflecting a broader move away from dashboards and query consoles toward conversational, task-executing assistants. This mirrors the rise of agentic interfaces elsewhere in software.
The fourth is that governance has expanded to cover agents themselves. Controlling who can read which table is no longer enough; teams now need to govern agent identity, behavior, and token spend. The market is split between open, integrable approaches and self-contained, proprietary ones, and that choice will shape platform decisions for years.
A fifth shift is emerging more unevenly: the long-standing separation between transactional (OLTP) and analytical (OLAP) systems is starting to soften, with real-time engines and embedded operational databases appearing inside analytical platforms. If that convergence holds, it could reshape data architecture well beyond any single vendor.
Databricks vs Snowflake: agentic data engineering in 2026
Dimension | Databricks - Data + AI Summit 2026 | Snowflake - Summit 2026 |
Posture | "The database your AI agents deserve" - depth & engineering control | "Bring agentic AI to all your data" - breadth & ease of adoption |
Context layer | Genie Ontology - self-improving context engine that learns what your data means | Horizon Context + Semantic Studio + Semantic View Autopilot - define shared business logic without deep SQL |
Interface | Genie suite - Genie Code, Genie One, Genie Agents, app builder | CoCo (ex-Cortex Code, now a native desktop app) + CoWork (ex-Snowflake Intelligence) |
Build-your-own agents | Agent Bricks - any model, any harness, isolated sandboxes (100k+ agents built, 1+ quadrillion tokens/yr) | No direct equivalent - focused on agents reaching data, not building them |
Governance philosophy | Open - Unity AI Gateway integrates outside security/identity vendors; Omnigent open-sourced | Closed - Agent Identity, Trust Center, proprietary prompt-injection detection |
Lake format | Apache Iceberg V3 (GA) | Apache Iceberg V3 (GA) via Horizon Catalog on Apache Polaris, bi-directional R/W |
OLTP/OLAP divide | Lakehouse//RT (Reyden engine) + Lakebase - moving to collapse it | No counteroffer today |
As of the 2026 summit season (June 2026).
What this means for data teams
For teams planning their next few years of data and AI work, a few implications follow directly.
Platform choice is becoming a foundation choice. The decision is less about storage or compute price and more about whose context engine, interface, and governance model your agents will depend on. Because semantic definitions are increasingly where the real investment goes, it is worth asking how portable those definitions are before committing.
Cost is now an architectural concern. Routing agent activity through a platform's managed services brings governance and convenience, and it also carries a per-token cost that scales with agent adoption. Modeling that economics early - while agent counts are small - avoids surprises later.
And it is worth designing for change at the OLTP/OLAP boundary. Architectures that assume transactional and analytical systems must remain separate may not age well if the convergence underway continues.
Where Upriver fits
The direction these summits confirmed is the one Upriver was built for. We believe agentic data engineering only works when the context layer comes first - when agents operate against a living, shared understanding of what an organization's data actually means - and when the engineer stays in control of direction and quality while agents handle the execution.
That is the foundation we are building: a platform where context is treated as the core asset, where data engineers shape and steer the work rather than hand-coding every step, and where AI does the heavy lifting without taking humans out of the loop. The largest platforms in the category have now validated the same thesis from their own angles. We think the teams that get the most out of this shift will be the ones that invest early in their context layer and keep their engineers firmly in the driver's seat.
If that is the future your team is planning for -> see what we're building at Upriver.
Frequently asked questions
What is agentic data engineering? Agentic data engineering is the practice of putting AI agents in the loop for the tasks data engineers have always owned - modeling data, building and maintaining pipelines, defining metrics, monitoring quality, and answering questions against governed data - while the engineer sets direction, reviews output, and stays accountable for what ships.
What did Databricks announce at Data + AI Summit 2026? The headline release was Genie Ontology, a self-improving context engine. Databricks also matured Agent Bricks into a full agent-development platform, shipped the Unity AI Gateway for agent governance, open-sourced the Omnigent framework, and pushed infrastructure with Lakehouse//RT, Lakebase, and GA of Apache Iceberg V3.
What did Snowflake announce at Summit 2026? Snowflake leaned into context and interface: Horizon Context with Semantic Studio and Semantic View Autopilot, the renamed CoWork and CoCo agent products, Cortex Training and Cortex Sense, an Agent Identity governance stack with the Trust Center, the Polaris-based Horizon Catalog with Iceberg V3, and Snowflake Datastream.
What is the difference between Genie Ontology and Horizon Context? Both are context-layer products that teach AI agents what your data means. Genie Ontology (Databricks) is a self-improving engine aimed at teams building their own agents; Horizon Context (Snowflake), with Semantic Studio, emphasizes defining shared business logic without deep SQL. Same bet, two philosophies - whoever owns the semantic layer owns the agent's trust.
Why does the context layer matter for AI agents? An AI agent is only as trustworthy as its understanding of what the data means. The context (semantic) layer gives agents the business meaning behind the schema - turning "the model can write SQL" into "the model writes SQL that's actually right for the business." In 2026 it shifted from optional tooling to foundational infrastructure.
