Anthropic Scion評測:一天3萬次合併,逼Git退場 | Anthropic Scion Review: 30K Merges/Day Force Out Git
By Kit 小克 | AI Tool Observer | 2026-10-21
🇹🇼 Anthropic Scion評測:一天3萬次合併,逼Git退場
Anthropic Scion 是 Anthropic 為了解決「AI 寫 code 太快、Git 跟不上」而自己打造的版本控制系統,直接取代了內部用了多年的 Git。這不是行銷話術——根據 Anthropic 官方工程部落格公開的數字,今年 5 月他們每天合併約 2,500 次程式碼變更,到 9 月已經飆到超過 3 萬次,整整翻了 12 倍,原因正是 Claude Code 這類 AI 代理大量自動寫 code、送出變更。
為什麼 Git 撐不住了
Anthropic 說得很直接:「寫 code 已經不是瓶頸,瓶頸是整套為人類規模設計的開發基礎設施——版控、build、CI。」他們的 monorepo 主幹歷史逼近 100GB,一份壓縮後的工作副本就超過 25GB,CI 每天要 checkout 程式碼大約 2,500 萬次(還沒算上開發環境、agent sandbox 跟研究用的 job)。如果照標準做法抓取,換算下來等於每天 600PB 以上的 Git 流量。團隊先試過淺層複製、稀疏簽出、promisor remote、把 snapshot 丟進 blob storage,甚至一度跑了數百個內部 Git mirror——全都只是拖延問題,沒解決問題。
Scion 到底是什麼
Scion 是一套以「change list」為核心的版本控制系統,概念上更接近 Google 內部的 Piper,而不是分散式的 Git。Anthropic 一支 7 人小組,花了 5 個月把整個 monorepo 搬過去,同時配套做出了程式碼審查工具 Antique。這套系統是為「數萬名工程師 + 成群 AI 代理同時寫 code」這種規模量身訂做的。
跟一般開發者有什麼關係
這裡最值得注意的是 Anthropic 自己講的老實話:Scion 不是給一般專案用的。官方工程部落格明白寫著,除非你的團隊也是數萬工程師等級的超大 monorepo,否則 Git 仍然是對的選擇。換句話說,這與其說是「你也該丟掉 Git」的宣傳,更像是一份誠實的規模化案例報告——AI 代理把程式碼產出量推到某個臨界點之後,連版本控制這種「早就解決的問題」都得重新發明。對 99% 團隊來說,真正該學的不是換工具,而是先搞清楚自己離那個臨界點還有多遠。
好不好用,試了才知道。
🇺🇸 Anthropic Scion Review: 30K Merges/Day Force Out Git
Anthropic Scion is the in-house version control system Anthropic built after discovering that Git couldn't keep up with how fast its own Claude Code agents write code. This isn't hype — per Anthropic's official engineering blog, the company merged about 2,500 code changes a day in May 2026. By September, that number passed 30,000 a day, a 12x jump driven almost entirely by AI agents generating and submitting code changes.
Why Git Couldn't Keep Up
Anthropic's own framing is blunt: "Coding is no longer the bottleneck — the bottleneck is all of the developer infrastructure built for human scale: source control, builds, and CI." Their monorepo's main branch history is nearly 100GB, a compressed working copy exceeds 25GB, and CI checks out code roughly 25 million times a day — before counting developer environments, agent sandboxes, and research jobs. Fetched the naive way, that works out to over 600PB of Git traffic daily. The team tried every standard workaround first — shallow clones, sparse checkouts, promisor remotes, blob-storage snapshots, even several hundred internal Git mirrors — but those only delayed the problem instead of solving it.
What Scion Actually Is
Scion is a change-list-based version control system, conceptually closer to Google's internal Piper than to distributed Git. A team of just seven engineers built it — along with a companion code review tool called Antique — and migrated Anthropic's entire monorepo over in five months. It's purpose-built for a scale most teams will never hit: tens of thousands of engineers plus swarms of AI agents writing code simultaneously.
What This Means for Everyone Else
The most useful part of this story is Anthropic's own honesty: Scion isn't for ordinary projects. Their engineering blog says it plainly — unless your team is also running a monorepo at tens-of-thousands-of-engineers scale, Git remains the right tool. So read this less as "you should ditch Git too" and more as an honest scaling case study: once AI agents push code output past a certain threshold, even a "solved problem" like version control needs reinventing. For the other 99% of teams, the lesson isn't to switch tools — it's to figure out how far you actually are from that threshold.
好不好用,試了才知道。
Sources / 資料來源
- Anthropic 官方工程部落格:Redesigning source control at Anthropic
- The Register: Anthropic reinvents source control as Claude floods the world with code
延伸閱讀 / Related Articles
- Tenstorrent評測:拒絕英特爾收購,改融資15億美元 | Tenstorrent Review: Turns Down Intel, Raises $1.5B
- RAM漲價評測:AI吃光記憶體,32GB飆破450美元 | RAM Price Surge Review: AI Eats Memory, Hits $450
- Claude破解NIST曲線評測:25年懸案,30分鐘解開 | Claude ECDSA Review: 25-Year NIST Mystery Solved in 30 Min
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
張貼留言