Key Notes
- Anthropic says daily merged code changes climbed from about 2,500 in May to more than 30,000 in September as Claude produced code at agent scale.
- Its seven-person team built Scion, a server-first source-control system, and migrated the monorepo away from Git in five months.
- Anthropic reports that Scion uses 200x less CPU and memory than its earlier Git-mirroring infrastructure and cuts CI and research checkouts from minutes to seconds, but does not plan a public release.
Anthropic says Claude is now producing code quickly enough to force a redesign of the infrastructure that stores, reviews and delivers it. The company’s engineering team built Scion, an in-house source-control system, after the daily number of merged code changes rose from roughly 2,500 in May to more than 30,000 by September. At that scale, Anthropic says source control—not writing software—had become one of the biggest constraints on development.
Scion is not a public replacement for Git. It is a system designed around Anthropic’s unusual workload: a giant monorepo, millions of daily checkouts and a growing fleet of coding agents that are online by default. The company fully migrated its monorepo away from Git in five months, while a seven-person team also built a review surface called Antique, code search and new access-control tooling. Anthropic’s engineering post presents the project as an early look at what source control may need to become when agents write code at a scale built neither for human teams nor ordinary CI pipelines.
Git was Built for a Different Kind of Client
Git’s decentralized design is a strength for the vast majority of software projects. A developer can hold a complete repository locally, work offline, inspect history and calculate differences without asking a central server. Anthropic found those same properties expensive when its clients were mostly agents and jobs that remained connected to the network and rarely needed the full repository.
The scale of the monorepo made the mismatch more visible. Anthropic says its main-branch history is nearing 100 GB, a compressed working copy exceeds 25 GB and its CI system alone performs about 25 million code checkouts per day. Fetching that activity naively would amount to more than 600 PB of Git traffic every day. The company had used shallow clones, sparse checkouts, mirrors and snapshot seeds, but maintaining hundreds of Git mirrors became harder as the merge rate accelerated.
Scion Moves Source Control to the Server
Scion puts the authoritative work on services instead of every client machine. The server can create commits, perform merges and rebases, answer history and blame questions, and calculate differences. Agents download only the directory trees and files needed for a task, while many jobs can skip a working copy entirely and call source-control APIs directly.
That approach changes the unit of work. Scion stores server-side changesets that record the files changed from a parent revision, along with content-addressed references to compressed files. The system can then derive data for history, blame, copy detection or search as needed. Anthropic says this avoids the need for a fully materialized copy of the repository, prevents a single machine from becoming the scaling limit and lets caches respond to immutable object requests without constantly coordinating with one another.
Faster Worktrees, Smaller Infrastructure
Anthropic says Scion’s CPU and memory footprint is 200x smaller than the Git-mirroring infrastructure it replaced. It also says CI and research jobs now check out code roughly 100x faster, moving from minutes to seconds, with CI alone gaining 25% of CPU capacity. Those are company-reported figures from an internal system, but they explain why the team chose a migration rather than another round of optimization around its existing Git setup.
The benefits matter especially for agents. A task that needs to compare revisions can ask Scion’s server for the diff instead of downloading a repository first. Sparse and virtualized checkouts materialize only the files a job reads. This removes a repeated cost from every agent sandbox, experiment and CI job—one that becomes consequential when the number of autonomous code-writing tasks rises into the tens of thousands each day.
Claude Helped Build the System that Claude Required
Anthropic describes a feedback loop at the center of the project. Claude’s ability to generate software helped create the source-control bottleneck, but it also helped a small team build the infrastructure to address it. The company says an early internal version of Claude rewrote a Git-to-Scion ingestion pipeline into a version that used 20x less memory and ran four times faster, while retaining key ordering and data-structure constraints.
That does not mean people stopped making engineering decisions. Anthropic says reviewers still examine every change request, but increasingly focus on intent, system constraints and architectural trade-offs instead of reviewing every line manually. It also describes cases where humans had to reject seemingly plausible agent-generated solutions, such as brute-force scans of a commit log where the correct answer was to query a review service. That division of labor echoes our reporting on Anthropic’s workforce data, which found agents logging far more runtime than their human handlers while engineers took on more oversight work.
Scion is a Signal, Not a Product Launch
Anthropic does not plan to release Scion as a general-purpose product. It says the system is too closely tied to its own code review, search and CI stack, and that its economics only make sense for exceptionally large monorepos or huge swarms of agents. Git remains the better model for most organizations, including those that need offline development or a portable toolchain.
Still, the company expects other organizations to encounter similar pressure points within the next six to 12 months as AI coding tools become more capable. The practical lesson is not that everyone should abandon Git. It is that agent-heavy teams may need more server-side APIs, more composable sparse checkouts and infrastructure designed for machines that can generate, test and merge code far faster than a human team can prepare local copies of a repository.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
