When Real-Time Meets Agents: Why We've Been on This Path All Along

A Shared Consensus Has Arrived

We're glad to see Databricks launch Lakehouse//RT.

This is a moment worth celebrating, not just for Databricks, but for the industry as a whole. When a player of this scale puts "real-time" front and center as a flagship capability, it sends a clear signal: real-time workloads are moving from the specialized systems of a few teams to a shared expectation across the entire industry.

Why now? Because AI agents are on the rise.

For most of their history, data systems served people. People can tolerate a few seconds of waiting; they can linger in front of a dashboard for a moment longer. Agents are different. Agents are always on, reasoning in loops, and unpredictable in their concurrency. Every decision they make depends on getting fresh, complete, trustworthy data the instant they need it. Latency that's merely "a little slow" for a human is often "can't function at all" for an agent.

That's why we believe real-time is the most important workload of the AI era. Real-time data ingestion and low-latency querying are becoming the bedrock of intelligent applications. With this launch, Databricks has helped turn that conviction into industry consensus.

Consensus is a starting point.

We've Been Cultivating This Path for Years

For us, this direction isn't something we just discovered. We've been working on it for years, and we've made it more complete, and more mature.

First, let's be clear about one thing: real-time has never been only about making queries faster.

True real-time is a complete loop: data can be ingested in seconds, it can be updated and modified in real time, and it can then be queried at very low latency under high concurrency. Ingestion, update, query. None of the three can be missing. Solve only one link, and what you get is "faster queries," not "true real-time."

This complete loop is something we've been running in production for years. It has been tested and refined by scale, by high concurrency, by real business workloads. It is mature, and it is proven. And real-time data updates in particular, making fast-changing data mutable and instantly effective, are the hardest and most often overlooked link in the real-time loop. They are exactly where we've invested for the long haul, and where we've become distinctive and leading in the industry.

It's worth noting that the newest entrant to this space is, for now, in Beta and supports read-only workloads only. This isn't meant as a criticism, quite the opposite. It underscores something important: getting real-time right, in a way that is both complete and mature, is a discipline that demands long-term focus. This isn't a question of how many features you have; it's a question of maturity and dedication. We're glad to walk this validated path alongside more and more people.

(We'll dive into the specific performance numbers and benchmarks in a dedicated follow-up article.)

Open and Autonomous

Databricks emphasizes open data formats, and we wholeheartedly agree with that direction. Freeing data from proprietary formats is the right way for the industry to evolve.

But "open" actually has two layers, and they're worth separating.

The first layer is openness of storage format: can you query directly on open formats like Iceberg and Delta Lake, without moving or copying the data? On this layer, the industry is rapidly converging.

The second layer is autonomy of platform and engine: do your query engine, your governance, your entire stack of capabilities have to run inside one specific vendor's platform? If they do, then even when your data format is open, you remain locked into a platform.

Our position is simple: both layers should be open. Our engine queries data directly on Iceberg and Delta Lake. No movement, no copies. At the same time, it isn't bound to any single cloud or platform; through BYOC, you can deploy it within your own environment. Open data, a free engine, and autonomous deployment. That is the complete meaning of no vendor lock-in.

From Customer-Facing to Agent-Facing

For the past decade, the primary battleground for real-time analytics has been customer-facing: high-concurrency user applications, live dashboards, personalized recommendations, the scenarios where every click expects an instant response. This is our heritage strength, and the place we've validated in production for years.

Now, data systems are welcoming an entirely new kind of consumer: agent-facing.

Agent-facing workloads are fundamentally different from customer-facing ones. Queries are no longer written and fixed by people; they're generated dynamically by models. Access patterns are unpredictable, and concurrency surges in bursts. What an agent needs isn't just the result of a single SQL statement, but an understanding of business semantics.

We've already extended our capabilities naturally from customer-facing to agent-facing, making us one of the few real-time platforms able to serve both people and agents as consumers.

But serving agents well takes far more than handing them "faster queries."

AgentBase: Building the AI Agent Infrastructure of the New Era

What agents need is a full set of infrastructure built natively for them.

That's why we're soon introducing AgentBase, an infrastructure product purpose-built for AI agents. It will support agents operating on top of enterprise data across four dimensions:

• Semantic: Help agents understand business semantics, so that every agent can correctly execute multi-step business-logic tasks, surface errors instead of failing silently, and reason through complex business problems more effectively.
• Context: Provide agents with managed, reusable, trustworthy context, so that every step of their reasoning stands on the right information.
• Observability: Make agent behavior, queries, and data access observable and traceable end to end. Agents should be visible and auditable inside the enterprise.
• Governance: Govern not just the data, but the agents themselves, their permission boundaries, what they can and cannot do. In an era where agents become the primary consumers of data, the object of governance must extend from data to the agent itself.

Giving agents access to data is only the beginning. Providing agents with a complete operational foundation is the infrastructure this era truly needs.

AgentBase is coming soon.

In Closing

Once again, thank you to Databricks for helping make "real-time" an industry consensus. Consensus is a starting point, not the finish line.

While the industry is still chasing "faster lakehouse queries," we've already traveled the road to a mature real-time loop, and we've set our sights on the next layer: defining native data and intelligence infrastructure for AI agents.

That's what we set out to do, to build the AI agent infrastructure of the new era.

The Next Step

To experience and reap the benefits as industry leaders have, consider trying PhoenixAI Cloud for free. This move could be the best decision for your engineers, analysts, and overall business, paving the way for unmatched flexibility, scalability, and cost-effectiveness. Get started here.

---