Anthropic Replaced Git With Its Own 'Scion' After AI Agents Broke It
Anthropic moved its entire codebase off Git onto a custom system called Scion, built by a seven-person team and rolled out in five months, after Claude's AI agents pushed daily merges from 2,500 to more than 30,000. Here is what Scion actually is, why Git could not keep up, and whether this really means anything for your own git workflow.
Anthropic has moved its main codebase off Git onto an in-house version control system called Scion, with a companion code-review tool called Antique. A team of seven built both and migrated in five months. The trigger was scale: as Claude's AI coding agents took over more of the work, daily merges jumped from about 2,500 in May 2026 to over 30,000 in September, overwhelming Anthropic's Git setup. Scion keeps no complete copy of the repository anywhere, not even on the server, and serves code over an API so agents and build systems rarely need to clone anything. Anthropic says it is "by no means a suitable replacement for git for general use" and is not releasing it. So no, your Git workflow is not in danger, but the reason Anthropic hit this wall is a sign of where AI-driven development is heading.
Git is the bedrock of modern software. Almost every developer alive uses it, it has seen off every challenger for nearly two decades, and the safe assumption has always been that it is not going anywhere. So when one of the most closely watched companies in tech quietly announces it has ripped Git out of its core workflow and replaced it with something homemade, it is worth understanding exactly what happened, and what it does not mean.
In an engineering post, Anthropic revealed it has migrated its monorepo off Git onto a custom system it calls Scion. The headline number is the one doing the rounds: its developers, increasingly AI agents rather than humans, now merge code so fast that Git itself buckled. Here is the honest version of the story.
What Anthropic actually did
Anthropic built two tools and switched to them. Scion is the new version control system, used through a client it calls scn. Antique is the code-review tool that sits on top. Per Anthropic, "a team of seven people built both Scion and Antique," and the company "fully migrated our monorepo from git in five months."
The reason, in Anthropic's own words: "In May 2026, Anthropic averaged around 2,500 code changes merged per day; as of September, we merged over 30,000 per day." That is more than a tenfold jump in four months. The cause is the obvious one for an AI company, its own models. As Claude-powered coding agents write and ship more of the code, the number of changes flying into the repository exploded past what a human-scale workflow was ever built to handle. This is the same shift behind the rise of AI "vibe coding", just turned up to an industrial volume inside a frontier lab.
Why Git could not keep up
Git was designed in 2005 for a specific picture of software work: many independent humans, often offline, each with a full copy of the project history on their own machine, occasionally syncing up. That distributed, offline-first design is exactly its genius, and exactly the wrong fit for a swarm of always-online AI agents hammering a single giant repository thousands of times an hour.
At Anthropic's scale the cracks showed as sheer overhead. The company says it was running several hundred Git mirrors at peak just to keep up, that its main-branch history had ballooned, and that it was doing tens of millions of CI checkouts a day. Every one of those agents cloning, fetching and pushing against a full repository is enormous duplicated effort. Git's model assumes the expensive thing, a complete local copy, is worth it because you rarely pay for it. Flip that assumption, and you are paying for it constantly.
How Scion works, in plain English
The core idea is a near-total inversion of Git. In Anthropic's words, with Scion "there is no fully materialized copy of the repo anywhere, even on the server." Instead of every client holding the whole history, code and history live in distributed storage and are served over an API, so an agent or a build job can query just the slice it needs without cloning anything.
The payoff Anthropic reports is dramatic, though these are the company's own figures and worth reading as such: infrastructure "200x smaller than our previous git mirroring infrastructure in terms of CPU and memory requirements," checkouts "two orders of magnitude faster (in seconds rather than minutes)," a "25% CPU capacity win" for CI, and cache hit rates "over 99%." In short, they stopped moving whole repositories around and started serving code like a database answers queries.
| Git | Scion | |
|---|---|---|
| Where the repo lives | A full copy on every machine | No complete copy anywhere, even the server |
| How clients get code | Clone and fetch the history | Query the slice they need over an API |
| Designed for | Distributed, offline human developers | Always-online AI agents at huge volume |
| Released to the public? | Open source, universal | No, internal to Anthropic |
Didn't Meta and Google already solve this?
Yes, largely, and Anthropic says so. The post credits Google's Piper and Meta's Sapling and its server component Mononoke, noting that Mononoke "pioneered many of the techniques we use in Scion, in particular the Bonsai data model." Giant-monorepo problems are not new; Google and Meta hit them years ago and built server-centric systems to cope.
That did not stop a row over credit. A former Meta engineer who worked on that design argued that Scion leans so heavily on Mononoke that Anthropic had not, at first, acknowledged it enough, framing the issue as "attribution, not copying." He does not allege Anthropic reused any code, and Anthropic updated its post on October 15 to expand its acknowledgement of the projects that pioneered these techniques. The fair reading is that Scion applies well-established big-tech methods to the new pressure of AI-agent volume, and the live dispute is about how fully that prior art was credited, not about stolen code.
So is Git dying?
This is where the hot takes get ahead of the facts. Anthropic is unusually blunt about it: Scion "is by no means a suitable replacement for git for general use." It is not releasing Scion as a product. The company even says "it's difficult to offer Scion in a form that would be broadly useful to other organizations," and frames the post as sharing design ideas, not shipping a Git-killer.
For the overwhelming majority of teams, who are not merging 30,000 agent-authored changes a day, Git is completely fine and remains the right default. What is genuinely true is narrower and more interesting: AI agents are the first thing in twenty years to stress Git's core assumptions at scale, and the companies feeling that pressure first are quietly rebuilding their plumbing. If you have felt Git strain under a big monorepo and wished for something faster, the live options are not Scion but tools like Jujutsu (jj), a modern, Git-compatible front end, or Meta's open-source Sapling. Git the format is not dying. The assumption that everyone needs a full local clone might be. For how the pressure of AI is reshaping other core infrastructure, see our pieces on AI in the Linux kernel and the npm supply chain.
Frequently asked questions
What is Anthropic's Scion?
Scion is a custom version control system Anthropic built to replace Git for its internal codebase, used through a client called scn and paired with a review tool named Antique. Unlike Git, it keeps no complete copy of the repository anywhere, even on the server, and serves code over an API so AI agents and build jobs can fetch just what they need without cloning.
Why did Anthropic replace Git?
Because AI coding agents pushed its daily merges from about 2,500 in May 2026 to more than 30,000 by September, overwhelming a Git setup that already needed several hundred mirrors. Git's distributed, offline-first design is a poor fit for a swarm of always-online agents hitting one huge repository constantly.
Is Scion open source? Can I use it?
No. Anthropic is not releasing Scion and says it is hard to offer in a form useful to other organisations. It is sharing the design ideas, not the software, so you cannot download or run Scion yourself.
How is Scion different from Git?
Git gives every machine a full copy of the repository and its history. Scion keeps no complete copy anywhere, not even server-side, and instead serves code and history over an API. That makes it far lighter at massive scale but tied to Anthropic's own infrastructure.
Does this mean Git is dying?
No. Anthropic itself says Scion is "by no means a suitable replacement for git for general use." For almost all teams Git remains the right choice. The story is a sign that AI-agent volume is starting to stress Git's assumptions at the very largest scale, not that Git is going away.
Did Anthropic copy Meta's Mononoke?
Anthropic credits Meta's Sapling and Mononoke (and Google's Piper) as prior art, saying Mononoke "pioneered many of the techniques we use in Scion, in particular the Bonsai data model," and does not claim to have reused any code. A former Meta engineer who worked on that design argued Anthropic had under-credited it, framing the dispute as one of attribution rather than copying; Anthropic expanded its acknowledgements in an October 15 update.
A maximum-severity zero-day hit the software MSPs use to run their clients' networks
N-able rushed out an emergency hotfix for CVE-2026-86218, a perfect-10 pre-authentication flaw in its N-central platform that its own customer notice says was exploited in the wild. Because managed service providers use N-central to reach thousands of downstream businesses, one flaw can cascade. Here is what to patch, and the honest picture of what is and is not known.
Hackers used Cursor's AI coding agent in a real ransomware campaign: what the Aurora case shows
Researchers documented a ransomware crew using Cursor's AI coding agent, powered by Claude, to help break into companies. It is a milestone, but the AI assisted a human, it did not attack on its own. Here is what happened, and how to defend against it.
The AI refused to steal the secrets, so researchers split the request into harmless pieces. It worked
Researchers at ASSET detailed GhostSplice, which makes AI coding agents leak local secrets by splitting a refused request across trusted channels, each fragment harmless alone. It roughly doubled compliance, from 42 to 82 percent, though Claude models mostly refused. A warning about how agents combine tool inputs.
'Vibe coding' let anyone build an app. Now the security bill is coming due
AI tools now let people describe an app in plain English and ship it without writing code, nicknamed 'vibe coding.' It is genuinely empowering, and it is producing a wave of live apps riddled with basic, avoidable security holes. Here is what the scans find and why AI-written code fails this way.
Meta just entered the AI coding-agent wars. The big two are now a big three
With Muse Code, Meta has launched its own AI agent that writes, tests and ships software from your terminal, going straight at Anthropic's Claude Code and OpenAI's Codex. Its headline trick is running many AI coders in parallel without them tripping over each other. Here is what it is, what's genuinely new, and the honest caveats.
Even Linux is embracing AI now, and drowning in it at the same time
Linus Torvalds has made it official: arguably the world's most important open-source project 'is not one of those anti-AI projects.' At the same time, the volunteers who maintain the Linux kernel are being buried under a flood of AI-generated bug reports and patches. Both things are true, and together they are the story of open source in 2026.