Databricks收購Electric評測：AI代理擁專屬Postgres資料庫 | Databricks Acquires Electric: Every AI Agent Gets Own DB

8月 13, 2026

By Kit 小克 | AI Tool Observer | 2026-08-13

🇹🇼 Databricks收購Electric評測：AI代理擁專屬Postgres資料庫

Databricks收購Electric：AI代理人也要有自己的資料庫

Databricks宣布收購新創公司Electric，這是繼去年以近10億美元收購Neon之後，Databricks在資料庫版圖的又一步棋。這次收購的重點是Electric旗下的PGlite——一款可以直接跑在瀏覽器分頁、應用程式或AI代理沙盒裡的WebAssembly版Postgres資料庫。過去12個月，PGlite的每週下載量從100萬暴增到1300萬，顯示開發者對「輕量級、可攜式資料庫」的需求正在爆發。

為什麼AI代理需要專屬資料庫？

傳統作法是所有AI代理共用一個雲端資料庫，每次讀寫都要打一次網路請求。當你同時跑幾十、幾百個AI代理（例如自動化測試、多代理協作系統）時，這種架構會變成效能瓶頸，代理也很容易讀到過期資料，或彼此覆寫對方的工作結果。

Databricks的解法是讓每個AI代理擁有自己的本地Postgres實例，透過PGlite在沙盒裡直接讀寫，不必等網路。背後再靠Electric的即時同步引擎，把每個代理的本地狀態同步回中央的Lakebase Postgres系統，確保多個代理不會用到過時資訊、也不會重複做同一件事。

開源承諾沒有變

值得注意的是，Electric先前開源的核心元件——同步引擎、PGlite、Durable Streams、TanStack DB——收購後全部維持開源。唯一收掉的是付費的Electric Cloud託管服務。Electric團隊將併入Neon，也就是Databricks去年收購、現在是Lakebase資料庫服務基礎的那家公司。

對開發者代表什麼

本地優先架構：AI代理不用每次讀寫都連外部資料庫，延遲大幅降低
開源生態不受影響：PGlite等工具仍可自由使用，不綁定Databricks
資料庫巨頭卡位AI代理基礎設施：這是繼向量資料庫、記憶體系統之後，下一個兵家必爭之地

這筆交易沒有公開金額，但反映出一個清楚的趨勢：當AI代理從單一助手演變成成群結隊的自動化系統，基礎設施廠商都在搶著定義「代理該怎麼存資料」這件事。好不好用，試了才知道。

🇺🇸 Databricks Acquires Electric: Every AI Agent Gets Own DB

Databricks Acquires Electric: Giving Every AI Agent Its Own Database

Databricks has acquired Electric, the startup behind PGlite — a WebAssembly build of Postgres small enough to run inside a browser tab, an app, or an AI agent sandbox instead of on a separate server. It is Databricks latest database move after acquiring serverless Postgres company Neon for roughly one billion dollars last year. PGlite weekly downloads jumped from 1 million to 13 million in just 12 months, a signal that developers are hungry for lightweight, portable databases.

Why AI agents need their own database

The standard setup has every AI agent share one central cloud database, with each read and write triggering a network round trip. That falls apart once you are running dozens or hundreds of agents at once — think multi-agent pipelines or automated testing swarms. Agents end up acting on stale data or stepping on each others work.

Databricks fix: give each agent its own local Postgres instance via PGlite, running inside its sandbox with no network wait. Electric real-time sync engine then keeps that local state in sync with a central Lakebase Postgres system, so agents do not duplicate work or act on outdated information.

The open-source pieces stay open

Notably, Electric previously open-sourced components — the sync engine, PGlite, Durable Streams, and TanStack DB — remain open source after the acquisition. The only casualty is the paid Electric Cloud hosted service, which is shutting down. The Electric team is joining Neon, the company Databricks acquired last year that now underpins its Lakebase database service.

What this means for developers

Local-first architecture: agents skip the network round trip for every read/write, cutting latency significantly
Open source stays intact: PGlite and friends remain free to use, not locked to Databricks
Database vendors are racing to own agent infrastructure: after vector databases and memory systems, how agents store data is the next battleground

Deal terms were not disclosed, but the move points to a clear trend: as AI agents shift from single assistants to swarms of automated systems, infrastructure vendors are all racing to define how agents should manage their own data. 好不好用，試了才知道。

Sources / 資料來源

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

搜尋此網誌

AI小貼士