When most code is written by agents, the systems built around human-paced development become the limit. This post reports that at Anthropic source control reached that limit in 2026, and describes the replacement: Scion, a centralized version control system designed on the assumption that its clients, mostly agents and CI jobs, are always online. It also gives the team's account of what building Scion with Claude was like.
The post was published on anthropic.com under the Engineering label on 13 October 2026 and updated on 15 October "to include more detail on the projects that pioneered many of the techniques we use in Scion" (the post's own notice). The earliest Wayback Machine capture is from 20 October 2026, after the update, so the original wording could not be compared. It is signed "Nick G. and the source control team" with seven further initials and first names. The post is the version of record; Scion is internal and has no public repository or paper. Every figure in it is an internal measurement that cannot be checked from outside. This note cites the post's sections and its six images by caption.
Terms used in this note
- Monorepo. A single repository holding most of an organization's code.
- Changeset. Scion's unit of work: an entry in a write-ahead log that records the files changed relative to the first parent and content-addressed pointers to their compressed contents in blob storage.
- Derived data. Everything Scion computes from changesets (blame, path history, skip lists over the commit graph), each type defined recursively from a changeset's parents.
- Sparse and virtualized checkouts. Working copies that contain only requested directories, or that materialize files only when read.
- Antique. Anthropic's code review tool, built by the same team.
What the post describes
Why git stopped fitting
The section "Growing our monorepo" gives the scale: the main branch's history is almost 100 GB and a compressed working copy over 25 GB; there are on the order of 25 million checkouts a day for CI alone, which the post says would be over 600 PB of git traffic a day if fetched naively. Standard workarounds (shallow and partial clones, sparse checkouts, internal mirrors, repository snapshots) had been used, at the peak with several hundred git mirrors, and became harder to run as the merge rate rose. The post's diagnosis is that the company used git as a centralized system with one authoritative copy while git's design serves offline, distributed work.
Design principles
The section "Principles for agent-scale source control" lists six principles, and names Google's Piper and Meta's Sapling as prior art; footnote 2 says Sapling's server, Mononoke, and its Bonsai data model "pioneered many of the techniques we use in Scion".
- Clients are always online and fetch data when needed.
- The server answers history and blame queries and creates commits, merges and rebases; the client keeps a bounded cache and lays out files.
- The same queries are exposed by API, so many jobs need no local repository (the post names automated bisect and revert pipelines).
- Every component scales horizontally; no fully materialized copy of the repository exists anywhere, and there are no packfiles to maintain.
- Commits are cheap, incremental and lock-free, costing in proportion to the change, not the repository; refs other than main and release branches are discouraged.
- Working copies are optional and sparse by design.
How Scion works
Changesets are created server-side; moving a ref is a compare-and-swap on a single Spanner row, separate from creating changesets (section "Data structures"). A small fleet of servers answers API calls and caches content-addressed data. For checkouts, a per-node daemon starts from a periodically refreshed snapshot and requests missing trees and blobs from regional cache servers, which fall back to the authoritative servers; all three tiers answer the same RPCs, and the tiering is enforced to isolate failures (section "Scaling Scion"; image "Scion's checkout architecture").
What it reports
All figures below are Anthropic's internal measurements, reported without a description of how they were collected or over what period, except where noted.
- Merge volume. The text says the company averaged around 2,500 merged changes a day in May 2026 and "as of September, we merged over 30,000 per day". The chart "Changes merged per day" (January to September 2026) labels a peak of 30,401 on 9 September; its 7-day average line ends near 18,000 (read from the chart, which prints no value for the average). The text's "over 30,000 per day" therefore matches the single-day peak rather than a typical day in September. The chart also shows the switch from pull requests to Antique change requests over June and July, and a dip labeled "Quiet week" at the end of July.
- Infrastructure. Each cache tier keeps a hit rate over 99%; the Scion footprint is "200x smaller" than the previous git mirroring infrastructure in CPU and memory; CI and research jobs check out code "two orders of magnitude faster (in seconds rather than minutes)"; CI gained 25% CPU capacity (section "Scaling Scion"). The command-time image below gives smaller median ratios for checkouts (8.0 times at the start of a compute job, 8.8 times in a CI job); the post does not say what comparison the "two orders of magnitude" refers to, for instance whether it includes the old mirror setup.
- Command times. The image "Everyday source control commands on the monorepo" gives median times, git against Scion's
scnclient. In developer environments: pull latest 87 s against 3.5 s, checkout 15 s against 1.1 s, rebase 13 s against 2.7 s, status 3.9 s against 19 ms, commit 1.4 s against 1.1 s, diff 407 ms against 67 ms. In agent environments: new worktree 7.1 min against 5.1 s, pull 3.0 min against 9.0 s, commit 18 s against 761 ms, status 4.3 s against under 0.2 s. Total source control time per change request in agent environments was 34 min against 3.7 min, labeled "≈11,000 agent hours a day". Checkout at the start of a compute job was 175 s against 22 s (labeled ≈85,000 machine hours a day) and in a CI job 31 s against 3.5 s (≈195,000 machine hours a day). The image, captioned "Observed performance improvements on native Scion repos", does not say how the hours were estimated or over which period the medians were taken. The CI figure is consistent with the text's 25 million CI checkouts a day: 25 million times the 27.5 s saved is about 191,000 hours (our arithmetic). - Build effort. A team of seven built Scion and Antique and "half a dozen related systems" and completed the migration in five months, while "the volume of code merged per day increased by 10x" (section "What we learned building Scion"). The chart "Changes to the Scion and Antique codebases per week" labels 12,257 changes since January.
- A rewrite by an early model. The first git-to-Scion ingestion pipeline was largely handwritten; an early internal snapshot of Claude Mythos Preview, asked to rewrite it "with no further instructions", produced a version using 20 times less memory and running 4 times faster while preserving the ordering and data-structure invariants. This is a single reported instance.
The team's account of working with Claude
The section "What we learned building Scion" gives the team's own observations, which are reports, not measurements:
- All engineers on the project "would now rate Claude as better at writing code than them".
- Engineers supplied system-level understanding and design tradeoffs, and "very often" their role was to tell Claude "no": agents sometimes broke core constraints, for example by reading the commit log linearly to find when a change merged instead of asking the code review tool.
- Adversarial review, used widely at Anthropic, could be "myopic"; one review's distrust of another internal service led Claude to add excessive validation, and engineers steered it back.
- Humans still review every change request, but examine "the intent of the change as opposed to reading every line of code."
- Claude produced hermetic simulations, fuzzing and rate limiting, and found optimizations such as one in ancestry queries that "missed one of our indexes".
- Claude instances in Slack channels, called "superfans", each watch the telemetry and logs of one component through Claude Tag; the team says they catch subtle issues that conventional alerts miss, without giving counts.
- With "hundreds or even thousands" of changes able to merge while tests run, the team dropped a global lock on the merge queue in favor of post-merge and pre-deploy testing, accepting that concurrent changes can conflict.
Scope and limitations
- Unverifiable figures. No figure in the post can be checked from outside, and none is given with a collection method, period or variance.
- Attribution of speed. The post credits Claude Mythos Preview and Claude Tag with accelerating the build, but gives no comparison with a team working without them; the five-month figure is set against the post's statement that such a system "would have taken a company years", which is an estimate.
- Not general advice. The post says Scion "is by no means a suitable replacement for git for general use", makes sense only for very large monorepos, and depends on spending tokens at a level few businesses would. It does not plan to release Scion, and instead proposes that git servers and forges adopt server-side RPCs, a different negotiation protocol and better-composing sparse and shallow features.
- Forecast. The expectation that "this class of problems" will become common "over the next six to 12 months" is the authors' prediction.
Research questions
- Does merge volume track delivered work? Comparing daily merged changes with measures such as deployed features, incident rates or reverts over the same months would show what the rise from about 2,500 to over 30,000 a day represents.
- How often do agents violate system constraints in practice? Counting rejected agent changes by cause across a migration would quantify the post's "very often".
- What does intent-level review miss? Sampling changes approved by intent review and auditing them line by line would estimate the defect rate of the review practice the post describes.
- Do monitoring "superfans" outperform alerts? Logging the issues each finds, with false-alarm rates, against conventional alerts over the same period would test the team's comparison.
- Can the server-side design be applied to git? Implementing the proposed server-side RPCs on a git forge and measuring checkout time and server load would test the post's recommendation outside Anthropic.
How it connects
- When AI builds itself reports other internal measures of Anthropic's own development, including the share of merged code written by Claude and code merged per engineer. This note connects them; the post does not cite the essay.
- How AI is transforming work at Anthropic studies how Anthropic engineers and researchers reported their work changing with Claude in August 2025, including a shift for some toward reviewing and revising Claude's code; the post's account is a later report from a single team. This note connects them.
- System Card: Claude Mythos Preview describes Claude Mythos Preview, the model the post credits with rewriting the ingestion pipeline.
References
- Nick G. and the source control team. (2026, October 13; updated October 15). Redesigning source control at Anthropic. Anthropic. https://www.anthropic.com/research/redesigning-source-control record
- Meta. Mononoke (Sapling server component), README. https://github.com/facebook/sapling/blob/main/eden/mononoke/README.md (cited by the post as prior art; not read for this note).