Databricks acquires Electric to embed Postgres databases in AI agent sandboxes

Databricks has acquired Electric, the creator of PGlite and a real-time database synchronization engine, to equip AI agents with their own embedded Postgres instances. The team will integrate with Neon, Databricks' serverless Postgres service acquired last year for roughly $1 billion, which now forms the core of its Lakebase database offering.

What Electric brings to the table#

PGlite is a complete Postgres implementation compiled to WebAssembly that runs directly in browsers, Node.js, or the isolated environments agents use to execute code safely. The project supports dynamic extension loading, including pgvector for vector embeddings—a critical capability for AI workloads. Downloads have surged from 1 million to 13 million weekly downloads over the past year.

More strategically significant is Electric's sync engine, which maintains a master Postgres database while keeping it synchronized near-instantaneously with browser interfaces, mobile applications, or agent runtimes. Think of it as bringing Figma's or Google Docs' collaborative multiplayer approach to databases and their agentic consumers. Building this kind of real-time synchronization from scratch is notoriously complex due to challenges like conflict resolution and partial replication, making it a logical acquisition target rather than build.

The founders James Arthur and Valter Balegas committed that existing open-source projects—the sync engine, PGlite, Durable Streams, and TanStack DB—will remain open. However, Electric Cloud, the hosted service, is being discontinued; existing users must self-host or migrate elsewhere.

Why agents demand different database patterns#

Traditional software distributes one central database across many clients, making the database the most durable component. AI agents flip this model upside down. According to Databricks' internal data, agents create approximately four times as many databases as human users on Lakebase, and typical projects now carry around 10 database branches with some exceeding 500 iterations.

Because Lakebase separates storage from computation and stores data in Postgres page formats on cloud object storage, spawning a new branch becomes a cheap copy-on-write metadata operation. A coding agent can bootstrap PGlite inside a sandbox, develop and validate against it locally, then either discard the instance or merge changes back upstream into a Lakebase branch—mirroring familiar code-branching workflows.

Remarkably, some workloads show average database compute lifespans under 10 seconds. As the Neon team notes, when agent-driven infrastructure reduces creation costs toward zero, "the number of applications explodes, and most of them are small." Traditional serverless databases still impose a floor on minimum viable app costs; Electric's lightweight architecture helps eliminate that constraint, unlocking an era of abundant, ephemeral database instances.