ClickHouse unifies its products into one platform, launching managed PostgreSQL, observability and security components, and an AI gateway.
ClickHouse acquires Xata, a PostgreSQL company whose branching lets agents snapshot production databases for testing without slowing production.
Andy Pavlo, ClickHouse's VP of Database Research, predicts agents will create more databases than humans by 2028, and text-to-SQL will replace BI dashboards.
ClickHouse Builds for a World Where Agents Outquery Humans
The story here is ClickHouse's bet that agents will hammer databases with volume rather than with harder questions, and that the winner will be whichever system keeps answering inside a tight time limit. Its move is to consolidate everything into one platform around its analytical core, adding managed Postgres for transactions and data movement plus security, observability and AI gateway products. Pavlo, who left a Carnegie Mellon professorship to run the company's new research lab, expects agents to out-create humans in databases within a few years.
Takeaways
Agents strain databases through volume, not complexity. Pavlo says agent queries are no harder than what people already write, but they arrive far faster. Agents loop query, result, query and run around the clock rather than on a human day-night pattern, which he expects to mean 10x to 1,000x more queries.
ClickHouse bets its core can absorb agent traffic. Pavlo says ClickHouse is uniquely positioned to answer that many queries within a tight time limit. The company has also reworked its SQL parser so that when an invalid query arrives, it throws an error back faster than it otherwise would.
Agents write invalid SQL more often than humans. Agents generate invalid SQL or reference tables that do not exist at a higher rate than people do, so fast errors with hints matter. They also use fewer SQL features than humans and have not learned the ClickHouse-specific ones that Postgres and DuckDB lack.
ClickHouse folds its products into one platform. The launch makes managed Postgres generally available and adds an AI gateway built on Langfuse, which runs prompt experiments and tracks cost. It also adds observability components that collect data from across an application stack, and a security product built on the earlier RunReveal acquisition.
Postgres becomes a front door inside ClickHouse. Users can query ClickHouse data through the Postgres front end and move data from Postgres into ClickHouse quickly, while a stock, fast Postgres handles transactions. ClickHouse is also buying Xata, whose branching lets agents snapshot production data to test without slowing it down.
A new research lab buys ClickHouse a longer horizon. Pavlo started in August, about three months before this conversation, to build ClickHouse Labs, which does exploratory research that may never reach the product and already includes early work on Postgres. As a Carnegie Mellon professor he had taught ClickHouse's published research to students.
What Pavlo predicts for databases
By 2028 agents will create more databases than humans, he says, and they will also create more branches of existing ones.
Text-to-SQL will eventually replace the BI dashboard, though dashboards stay useful for humans to comprehend things and can be paired with a tool that draws the pictures.
People could ship a database written by agents now, he says, but it needs guardrails and verification first.
PostgreSQL will still be the default database in five years, he predicts, the way Linux is the default operating system.
Is computer science still worth studying
Pavlo says a 20-year-old should still major in computer science and specialize in databases, because the discipline does not go away just because agents exist.
He expects the field to trend back to its roots in applied logic, philosophy and mathematics, and says the days of writing a pile of JavaScript and calling yourself a computer science major are gone.
Host Ruby Justice Thelot, who majored in computer science and graduated the year ChatGPT launched, says training your brain to think in a puzzle-and-logic way stays valuable however AI evolves, even if those skills stop being marketable.
