跳到主要內容

Scion評測:Anthropic自建版控,拋棄Git扛3萬次合併 | Scion Review: Anthropic Ditches Git for 30K Daily Merges

By Kit 小克 | AI Tool Observer | 2026-10-14

🇹🇼 Scion評測:Anthropic自建版控,拋棄Git扛3萬次合併

Anthropic 這週公開了一個挺反直覺的消息:為了因應 Claude 寫程式碼的速度,他們已經把整個 monorepo 從 Git 搬到自建的版本控制系統 Scion。原因很直接——今年 5 月,Anthropic 內部每天大概合併 2,500 次程式碼變更;到 9 月,這個數字衝到超過 30,000 次,成長超過十倍。Git 當初的設計假設是「每個 client 都要能離線跑完整副本」,這個假設在人類工程師身上成立,但在一大群 24 小時不睡覺、瘋狂開分支寫 code 的 AI coding agent 身上完全破功。

Scion 解決了什麼問題?

Scion 的核心設計哲學是「server 永遠在線,client 按需抓資料」。簡單說:

  • Server 端負責重活:歷史查詢、blame、建立 commit、merge、rebase 都在伺服器端完成
  • Client 端只留必要快取:本地維護一個有上限的物件快取,而不是整包倉庫
  • 效能數字很誇張:Anthropic 宣稱新架構的 CPU/記憶體用量比舊的 Git mirror 方案少了 200 倍,CI checkout 從分鐘級降到秒級

整個遷移由 7 名工程師、花了大約 5 個月完成,過程中甚至讓 Claude 自己改寫了一段 ingestion pipeline,把記憶體用量砍到 1/20、速度提升 4 倍——某種程度上是 AI 自己參與了淘汰 Git 的工程。

一般團隊要跟著拋棄 Git 嗎?

先說結論:不需要。Anthropic 很明確表示 Scion 跟內部基礎設施綁得太緊,不會開源,也沒有對外釋出的計畫。這不是一個「比 Git 更好的工具」,而是一個解決「單一團隊每天 3 萬次合併」這種極端規模問題的客製方案。如果你的團隊一天合併次數是兩位數而不是五位數,Git 搭配好的分支策略仍然夠用。

但 Scion 這個案例值得開發者留意的,是它揭露了一個趨勢:當 AI coding agent 的數量和寫 code 速度超過人類團隊規模的量級,原本為「人類協作」設計的工具鏈(Git、CI、code review 流程)都可能需要重新思考。這不是 Scion 獨有的問題,未來很可能會有更多團隊遇到類似的合併瓶頸。

好不好用,試了才知道。


🇺🇸 Scion Review: Anthropic Ditches Git for 30K Daily Merges

Scion is the new version-control system Anthropic built to replace Git inside its own monorepo — and the reason is almost absurd: AI coding agents made Git too slow to keep up. In May 2026, Anthropic merged about 2,500 code changes a day. By September, that number passed 30,000 merges a day, a more than 10x jump driven almost entirely by Claude-powered coding agents running around the clock.

What Problem Does Scion Actually Solve?

Git assumes every client needs a full, offline-capable copy of the repository — a fine assumption for human engineers, a terrible one when most of your "clients" are agents and CI jobs hammering the repo nonstop. Scion flips that model: the server stays authoritative and always reachable, while clients fetch only what they need.

  • Server-side heavy lifting: history queries, blame, commits, merges, and rebases all happen on the server
  • Bounded local cache: clients keep a capped object cache instead of a full repo clone
  • Dramatic efficiency gains: Anthropic claims roughly 200x lower CPU/memory footprint than its old Git-mirroring setup, with CI checkouts dropping from minutes to seconds

Seven engineers built Scion in about five months — and in a nice bit of recursion, an early Claude model rewrote part of the ingestion pipeline during development, cutting memory use 20x and quadrupling throughput.

Should Your Team Ditch Git Too?

No — and Anthropic isn't suggesting you should. Scion is tightly coupled to Anthropic's internal infrastructure, has no public code release, and isn't being open-sourced. This isn't a "better Git" for everyone; it's a bespoke fix for a scale problem most teams will never hit. If your merge volume is in the dozens rather than the tens of thousands per day, Git with a sane branching strategy is still the right call.

What's worth watching is the underlying trend Scion exposes: once AI coding agents outpace human-scale commit volume, tooling built for human collaboration — Git, CI pipelines, code review — starts cracking at the seams. Scion won't be the last team to hit this wall.

The only way to know if it's useful is to try it yourself.

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

留言

這個網誌中的熱門文章

Google Ironwood TPU v7 推理專用晶片解析:效能追平 NVIDIA、成本低 44%,AI 晶片戰爭正式開打 | Google Ironwood TPU v7 Explained: Matching NVIDIA Performance at 44% Lower Cost — The AI Chip War Heats Up

Claude Code 實測:AI 幫你寫程式到底行不行? | Claude Code Review: Can AI Really Code for You?

Cursor vs GitHub Copilot vs Claude Code:AI 程式助手大比拼 | AI Coding Assistants Compared: Cursor vs GitHub Copilot vs Claude Code