Databricks Bought a Database That Fits Inside an Agent

Databricks announced on August 11, 2026 that it has acquired Electric, the company behind PGlite and the Electric sync engine. Terms were not disclosed. Electric had raised roughly $7 million from a group of angels and funds including Spark Capital, Lunar Ventures, StepFunction, Firestreak, and Ninja Capital.

Small company, small round, and a genuinely consequential acquisition — because of what PGlite is.

Postgres, compiled to WebAssembly

PGlite is a full Postgres build that runs inside your application rather than on a server. A browser tab. A user's device. A Node process. An AI agent's sandbox. No connection string, no network hop, no separately provisioned instance.

That sounds like a novelty until you consider the numbers: weekly downloads went from 1 million to 13 million between August 2025 and August 2026. Thirteen-fold growth in a year is not a novelty curve. It is a curve driven by a use case that materialized suddenly, and the use case is agents.

Why agents broke the database model

Run one agent and a shared database is fine. Run a thousand concurrent agents, each executing a multi-step task with intermediate state, and the model falls apart in specific, expensive ways.

Every agent needs somewhere to put working state — retrieved documents, partial results, a scratch table, the record of what it has already tried. Handing all of them connections to a shared Postgres means connection-pool exhaustion, lock contention between agents that have nothing to do with each other, and a blast radius where one runaway agent's query degrades every other agent in the fleet. Provisioning a real database per agent means paying provisioning latency and cost on a resource that lives for ninety seconds.

The other option is what most agent frameworks actually do: keep state in memory as unstructured objects, and lose the ability to query it, join it, transact over it, or reason about it with anything other than more model calls. That is a large amount of engineering effort spent badly reimplementing a database.

PGlite gives each agent a real Postgres — transactions, SQL, indexes, the whole contract — that starts in milliseconds and costs nothing to provision because it is just a library in the process.

The sync engine is the other half

An isolated database per agent solves contention and creates a new problem: a thousand agents each holding a private view of the world, drifting apart.

Electric's sync engine is the answer, and it is the part Databricks most plainly wanted. Under the architecture Databricks describes, each agent runs a local PGlite instance and synchronizes state back to a central Lakebase Postgres. Agents get local-first speed and isolation; the platform keeps a coherent central record. Agents avoid acting on stale data or duplicating work another agent already finished.

That is a real distributed-systems primitive, and it is exactly the piece that has been missing from every multi-agent framework shipped so far. Orchestration frameworks are abundant. Shared, consistent, queryable state across a fleet of concurrent agents is not.

What happens to the open source

Databricks says PGlite, the sync engine, Durable Streams, and TanStack DB all remain open source. The one casualty is Electric Cloud, the hosted service, which shuts down.

That is the honest and expected shape of this kind of deal, and it is worth stating plainly: the hosted business was never the point. Electric raised $7 million and built libraries that 13 million weekly downloads depend on. Databricks did not buy revenue. It bought a team and a position in a layer that is about to matter enormously — and killing the open source would destroy the only thing that made the position valuable.

TanStack DB staying open is the detail front-end developers will care about most, since it pulls a large adjacent ecosystem into Databricks' orbit without requiring anyone to adopt Databricks.

The strategic read

Databricks has spent the last two years positioning around Lakebase — Postgres as an operational layer sitting next to the analytical lakehouse. The pitch is that AI applications need both: fast transactional state for the application and deep analytical context for the intelligence, without an ETL pipeline between them.

Electric completes that pitch at the edge. The stack now runs from a database inside an agent's sandbox, through a sync engine, into a central operational Postgres, into the lakehouse. Every layer is Postgres-compatible. Every layer is queryable with the same language.

Compare that to what a competitor has to assemble: an agent framework, a vector store, a key-value cache for agent state, an operational database, an analytical warehouse, and pipelines connecting all six. Simplicity is a real product advantage when the alternative is six vendors.

The thing to watch

The bet embedded in this acquisition is that agents become numerous rather than powerful — that the dominant deployment shape is thousands of small, cheap, concurrent agents doing narrow tasks, not a handful of expensive ones doing everything.

If that is right, per-agent state infrastructure is a foundational layer and Databricks just bought it for the price of a Series A. If the industry instead converges on fewer, larger, longer-running agents with big context windows and less need for external state, PGlite remains an excellent library serving a smaller problem.

Thirteen million weekly downloads suggests the market has already voted on which way this goes.