Anthropic has stopped treating code writing as the hard part. In an engineering post published on 13 October, the company says it averaged around 2,500 code changes merged per day in May 2026. By September it was merging over 30,000 a day. The constraint, in its account, moved downstream: source control, builds and continuous integration, all built for human pace, became the bottleneck.
The response was not a bigger Git setup. Anthropic says a team of seven built a new version control system called Scion, a code review surface called Antique, and related code search, access-control and code-landing systems, then migrated the whole monorepo off Git in five months, while merge volume was still climbing.
The arithmetic that forced the decision
The scale figures in the post explain why incremental fixes ran out. Anthropic says the history of its main branch alone is almost 100 GB, and a compressed working copy is over 25 GB. It performs on the order of 25 million code checkouts per day for CI, before developer environments, agent sandboxes and research jobs are counted. At its peak, it says, it operated several hundred internal Git mirrors to carry that load, and the mirrors got harder to run as the merge rate rose.
Git's core assumption is the problem Anthropic names: every clone is a complete copy that can work offline. Most of Anthropic's clients are not people on laptops. They are agents and automated jobs that are online by definition, and most of them need one directory or one diff, not the full history of the company.
What Scion does differently
Scion inverts that assumption. The server owns history, blame, commits, merges and rebases. The client keeps a bounded local cache and the working files it actually needs. Checkouts are sparse by default, containing only requested directories, and some jobs skip the checkout altogether: a CI job that needs the diff between a commit and its merge base can ask the server for it directly. Anthropic says it could not effectively reproduce a conventional global-lock merge queue at its scale, so it invested in post-merge and pre-deploy automated testing. For Scion and its related systems, human engineers still review every change request, but focus on the intended change rather than reading every line.
The process detail matters as much as the storage design. Anthropic is candid that its engineers spent much of their time telling the model no: agents, in its account, sometimes broke core system constraints to satisfy a pattern, and adversarial review could push the model into adding excessive validation. Human review did not disappear. It moved up a level, from lines to intent, which is a different skill and a different failure mode.
Read the gains with their source attached
Anthropic reports a striking set of results: an infrastructure footprint 200 times smaller in CPU and memory than its previous Git mirroring infrastructure, checkouts that take seconds rather than minutes — two orders of magnitude faster, in Anthropic's own characterization (company-reported) — and a 25% CPU capacity gain for CI alone. It also says an early Claude Mythos Preview snapshot rewrote its ingestion pipeline to use 20x less memory and run 4x faster, according to Anthropic's internal account.
Every one of those figures is Anthropic's own measurement, on its own systems, in a post written to explain its own project. Nothing in the post lets an outside team reproduce the comparison, and Anthropic has not released Scion for anyone to test. That does not make the numbers false. It makes them a vendor benchmark with an unusually detailed methods section, and they should be carried with that label attached.
Anthropic, to its credit, attaches the biggest caveat itself. It says Scion is by no means a suitable replacement for Git for general use, that the ideas only make sense for truly enormous monorepos, and that its internal components are too bespoke to offer in a broadly useful public form. Its stated plan is narrower: share the thinking with GitHub, GitLab and the open-source community, and argue that Git servers can adopt server-side operations without abandoning Git itself.
A practical angle for teams that are not Anthropic
Almost no team merges 30,000 changes a day, and almost no team should build a version control system. The transferable test is earlier in the chain. Count how many agent-written changes your review, CI and merge steps can absorb per day, and compare it with how many your agents can now open. If the second number is growing faster, your bottleneck has already moved, even if it is still measured in dozens rather than thousands. None of this is because Git lacks the relevant features: Git already has shallow-clone, sparse-checkout and promisor-remote capabilities. Anthropic's argument is that those mechanisms became insufficient at its particular scale and operating pattern, not that Git lacks partial-checkout functionality. The cheaper fixes come first: sparse and partial checkouts, server-side diffs for CI jobs that read one service, and review rules that judge intent on low-risk changes instead of queuing every line behind the same gate Anthropic just walked away from.
Sources
- [1] Anthropic Engineering, "Redesigning source control at Anthropic", published 13 October 2026 — primary source for all figures, system design, team size, timeline and the not-a-general-replacement caveatRead source
- [2] Daily AI Brief, "Anthropic Replaces Git With Scion for Its Monorepo" — secondary summary corroborating the sparse/virtualized checkout description and reported gainsRead source
- [3] Particle News, "Anthropic Moves Main Repository From Git to Scion" — secondary corroboration of the five-month migration, seven-person team and plans to share the approach with GitHub/GitLab/OSSRead source