Anthropic builds Scion to replace Git: 30,000 merges per day, a VCS infrastructure reduced by 200x
🔎 A merge every three seconds
Between May and September 2025, Anthropic's monorepo went from 2,500 to more than 30,000 merges per day. Do the math: a merge every three seconds, day and night, weekends included. At that pace, Git — the tool designed in 2005 by Linus Torvalds for human developers — stopped keeping up.
Anthropic's response was not to optimize, but to rebuild. Seven engineers, five months, and two tools built from scratch: Scion, a version control system that replaces Git internally, and Antique, the accompanying review tool. The figures announced on Anthropic's engineering blog are dizzying: checkouts in seconds, a VCS footprint divided by 200, and a cache hit rate above 99% on 100 GB of history.
Beyond the engineering feat, these numbers tell a bigger story: the scale of agentic workflows has outgrown what Git can absorb. And this is probably just the beginning.
The essentials
- Daily merges on Anthropic's monorepo went from 2,500 (May 2025) to more than 30,000 (September 2025): a 12x increase in four months, driven by coding agents.
- Seven engineers built Scion (the VCS) and Antique (the review tool) in five months.
- Announced results: checkouts in seconds, a VCS footprint divided by 200, and a cache hit rate above 99% on 100 GB of history.
- Scion is neither open source nor a product: it's a signal. Agentic throughput has outgrown tooling designed for humans.
- For the vast majority of teams, Git remains the right choice — but the breaking points are approaching faster than expected.
Recommended Tools
| Tool | Primary use | Price | Ideal for |
|---|---|---|---|
| Scion | Anthropic's internal VCS | Internal (not available) | Monorepos with very high agentic throughput |
| Antique | Code review integrated with Scion | Internal (not available) | Triage of agent-generated changes |
| Git | Standard distributed VCS | Free (open source) | Nearly all projects, all sizes |
| Sapling | Meta's VCS, monorepo-oriented | Free (open source) | Large organizations with massive repositories |
| Jujutsu | Modern Git-compatible VCS | Free (open source) | Developers seeking a healthier workflow |
| VFS for Git | File system virtualization for Git | Free (open source, legacy project) | Giant monorepos (Microsoft/Windows case) |
| Graphite | PR stacking and review | Free, paid team plan (November 2025, check graphite.dev) | Teams that merge frequently and often |
2,500 → 30,000 merges per day: it's throughput, not size, that broke
What brought Git to its knees at Anthropic wasn't an oversized repository. It was a flow that was too fast.
Between May and September 2025, the number of daily merges grew twelvefold. The cause is documented: agents — Claude Code chief among them — chain together parallel sessions that produce branches, commits, and changes at a pace no human team will ever come close to. Current models like Claude Opus 4.7 (Adaptive) or GPT-5.5 sustain long tasks from end to end, which multiplies the number of completed workstreams per session accordingly.
The consequences cascade: saturated merge queues, chained rebases, conflicts multiplying as branches cross paths, CI overwhelmed. Every link in the system, designed for a few hundred operations per day, buckles at a few tens of thousands.
One detail deserves to be highlighted: these figures come from Anthropic itself, which has every interest in showing its agents running at full capacity. Take them seriously, yes. Take them at face value without any critical distance, no. That said, the order of magnitude — one merge every three seconds — is consistent with what other heavily agentified companies report behind the scenes.
Scion: a VCS designed for throughput, not for local clones
Scion is Anthropic's internal version control system, built in five months by a team of seven engineers to replace Git on the monorepo. According to the engineering post, it relies on a centralized model: a single authoritative server, with clients kept lightweight and syncing incrementally.
The three key numbers
- Checkouts in seconds: fetching a working state no longer takes minutes. For thousands of agentic sessions starting continuously, this is the absolute prerequisite.
- VCS footprint divided by 200: disk cost and storage load collapse. A factor of 200 isn't an optimization, it's a category change.
- Cache hit rate above 99% over 100 GB of history: nearly all traffic is served without recomputation or retransmission. The system breathes instead of suffocating.
This design didn't come out of nowhere. It follows in the lineage of the centralized VCS built by tech giants for their monorepos — Google with Piper, described in Communications of the ACM in 2016, and Meta with Sapling. The difference lies in the driving constraint: at Google and Meta, the goal was to absorb tens of thousands of human developers. At Anthropic, it's to absorb tens of thousands of agents.
Important caveat: Scion is not open source, and no independent benchmark is available to date. The figures are Anthropic's own, presented without external audit.
Antique: because a fast VCS is useless if review collapses
Antique is the code review tool built in parallel with Scion — and it's perhaps the most interesting half of the project.
At 30,000 merges per day, traditional human review is mathematically impossible. Nobody reads 30,000 diffs a day. Antique therefore runs on a different logic: agents produce and pre-qualify, humans validate the changes that matter. Review becomes assisted triage — automatic summaries, grouped diffs, alerts on sensitive areas — rather than exhaustive reading.
It's the same trade-off as everywhere else in AI applied to work: automate the volume without ejecting humans from the decisions that carry real stakes. You find exactly this tension in Avatar AI for customer service: replacing without losing the human — replace mass processing, keep humans on the cases that matter.
My take: it's this part — review — that existing tools underestimate the most. GitHub, GitLab and the like have optimized human-to-human review. The agent-to-human flow is a different problem, and almost nobody has tackled it at this scale.
Git, born from a human crisis in 2005, stumbles on a machine crisis in 2025
Git hasn't aged badly. It does exactly what it was designed to do — and that's precisely the problem.
April 2005: Linus Torvalds wrote Git in a few weeks to replace BitKeeper and manage the development of the Linux kernel. The problem to be solved: thousands of human developers, distributed, working on the same code. The distributed model — everyone clones the full history — was the brilliant answer to that particular problem.
Twenty years later, the load profile has changed. Precedents had already shown this: Google built Piper for its giant monorepo (2016), Microsoft created VFS for Git in 2017 to virtualize the file system on the Windows repository, Meta open-sourced Sapling in November 2022 after years of internal use. Each time, the same finding: at very large scale, Git's distributed model costs more than it delivers.
The novelty of 2025 is the nature of the load. It's no longer humans in excessive numbers, it's agents in absurd numbers. And the irony doesn't stop there: Linus Torvalds himself now publishes Linux kernel patches written by Claude Code, as reported in Linus Torvalds publishes Linux kernel patches written by Claude Code. The tool born to coordinate humans sees its flow overrun by machines — including in its original project.
My verdict: Git is not dying. It is simply reaching the edge of its comfort zone. What seeks to replace it is not a commercial competitor, it's a change in usage regime.
Should your team care? Yes — but don't act yet
No, you don't need Scion. Yes, you should be watching your metrics.
If your team merges a few dozen PRs a day, Git is perfect and will stay that way for a long time. Anthropic's problems begin where agentic throughput becomes the bulk of the traffic — a point most teams haven't reached yet, but one that's approaching faster than anyone anticipated.
The warning signs to watch for
- Your fetch, rebase, and checkout are taking a few extra seconds, month after month.
- Your CI is queuing even though nothing has changed on the infra side.
- Your review backlog keeps growing while headcount stays flat — a sign that agents are producing faster than you can triage.
Three actions are worth more than a migration: measure the agentic share of your commits now; adopt PR stacking to make merges smoother; automate commits and reviews with AI, as detailed in Git + AI: automate your commits and reviews. And if you're looking for the right agents to equip your developers, our selection of best AI tools for code is updated regularly.
On a personal level, Jujutsu is worth a try: compatible with your existing Git repositories, it modernizes your workflow without a painful migration.
The real signal: development infrastructure is aligning with AI
Scion is a symptom before it is a product. It's the first public case of a company rebuilding its fundamental development tooling — versioning, one of the lowest building blocks in the stack — because its agents produce more than its humans.
Put this fact in series with Anthropic's other moves and the picture becomes clear. The company is raising record amounts to fund the agentic era, as we detailed in Anthropic raises $65 billion in Series H and launches Claude Opus 4.8: the day everything changed. It is turning its coding agent into a platform open to extensions, which we covered in Claude Code MODS: Anthropic turns its coding agent into a platform. And it is building in-house the infrastructure that this platform requires.
My take: we are witnessing the emergence of an "agent-first" infrastructure category. Likely next targets: CI, designed for thousands of parallel runs per minute; code review, already underway with Antique; and code storage and search, designed for machine consumers rather than human readers.
One bias to keep in mind: every announcement of this kind also serves Anthropic's commercial narrative — "our agents run at a scale that breaks the tools." That's probably true, and it's above all good for business. The two are not mutually exclusive.
Should you abandon Git? No — here's our verdict
No, and it's not even up for debate today.
Git remains the best default choice: an unmatched ecosystem, hosting everywhere, skills everywhere, tools everywhere. The alternatives each have a specific use case — Sapling for very large monorepos, Jujutsu for individual comfort, Graphite for stacking — but none of them justifies a migration on principle.
What you need to do is keep an eye on two things. First, your own agentic throughput metrics: that's the only indicator that truly matters. Then the ecosystem: if Anthropic ever opens up Scion, the equation will change completely. In the meantime, this kind of tooling will first emerge on the side of recent AI tools, which we track as they come.
❌ Common mistakes
Mistake 1: Reducing Scion to "a faster Git"
It's a different model: an authoritative central server versus full local distribution. The trade-offs aren't the same — offline work, forks, tooling ecosystem. The solution: understand the architecture before dreaming of a migration.
Mistake 2: Believing repository size is the problem
The variable that breaks things is throughput. A 50 MB repository with 30,000 merges per day blows up too — merge queues, cross conflicts, saturated CI. Watch your merges per day, not just your gigabytes.
Mistake 3: Looking for Scion on GitHub
Neither Scion nor Antique is open source as of today. The only reliable source remains Anthropic's engineering post. Beware of opportunistic projects squatting the name.
Mistake 4: Waiting until you're out of air to measure
The 12x factor in four months can hit any team that wires up agents in series. Instrument now: share of agentic commits, average merge time, CI queue depth.
❓ Frequently Asked Questions
Does Scion replace Git everywhere at Anthropic?
According to the engineering post, Scion replaces Git on the main monorepo, with a gradual migration. Nothing suggests any intention to replace Git for open source projects or external workflows, which typically remain on conventional platforms. Scion is an internal tool first and foremost — not a product.
Are Scion and Antique open source?
No, not to date. Anthropic has only published an engineering article detailing the architecture and the results. Precedents (Piper at Google, Sapling at Meta) show that full open-sourcing is never a given. Keep an eye on Anthropic's engineering blog for any announcements.
Why not optimize Git instead of replacing it?
A fair question. Band-aid fixes exist: partial clone, VFS for Git, beefed-up merge queues. Anthropic judged — and the post suggests — that agentic throughput called for a different model, not a patch. A debatable call for a fifty-person team, defensible at 30,000 merges per day.
What is the connection with Claude Code?
Direct: it's the agent sessions, Claude Code above all, that took merges from 2,500 to 30,000 per day. Every session generates branches, commits, and changes. Scion is thus, in effect, an infrastructure built so that Claude Code can run at full capacity without overwhelming the rest.
Should you migrate to Jujutsu or Sapling now?
Not out of panic, yes out of comfort. Jujutsu brings a healthier workflow at the individual level while staying compatible with your Git repositories. Sapling targets very large organization-wide monorepos. If your pain comes from agents, start by measuring and automating before switching VCS.
✅ Conclusion
Scion doesn't replace Git everywhere: it replaces Git where agents have taken control of throughput — and that tipping point is getting closer for everyone. To see where the rest of the ecosystem is heading, keep an eye on our current AI trends.