Databricks just made a bet that the future of AI infrastructure looks less like one big shared database and more like thousands of tiny, disposable ones. The company announced it is acquiring Electric, the startup behind the WASM-based Postgres project PGlite and the Electric sync engine, as agentic applications change how developers use databases. The deal is a direct response to a pattern Databricks is already seeing in production: agents don't use databases the way humans do.

Agents broke the database model

Traditional applications are predictable. They have known query patterns, fixed data access, and one managed Postgres instance serves them well. Agents are different in three fundamental ways:

• They decide what data they need at runtime, updating their context several times per second. That inner loop wants data in the same process.
• They run inside sandboxes, where a remote database connection adds latency that compounds across every reasoning step.
• They work in parallel groups, requiring both fast local context and a shared, current view of what other agents have done to avoid duplicating work or acting on stale state.

The numbers from Databricks' own platform back this up. Agents now create roughly four times more databases than human users do on Lakebase. The average project carries about 10 database branches, some run more than 500 branch iterations deep, and for certain workloads, the average database compute session is alive for under 10 seconds. The old model of one durable server per app simply does not map to this world.

What Electric actually built

PGlite gives every agent its own lightweight Postgres right where it runs, providing ultra-low latency access to local context. It has grown from 1M to 13M weekly downloads in just twelve months. The key technical trick: PGlite is a lightweight WASM Postgres build, packaged into a TypeScript library for the browser, Node.js, Bun, and Deno, with no need to install any other dependencies. And at under 3MB gzipped, it is small enough to live inside the agent sandbox itself rather than on a separate server.