Anthropic Builds Scion Source Control System to Replace Git, Merging Code Changes About 30,000 Times Daily
Anthropic disclosed in its engineering blog that it has spent five months building an internal version control system called Scion, completing a full migration from Git's single repository. The company stated that coding is no longer a bottleneck, as the bottleneck has shifted to the development infrastructure designed for human scale.
In May 2026, Anthropic merged code changes about 2,500 times daily; by September, this number exceeded 30,000. The main branch history is nearly 100GB, with compressed working copies over 25GB. Continuous integration jobs execute about 25 million code checkouts daily, which would generate over 600PB of Git traffic if pulled in the traditional way. The company previously relied on partial copies, snapshots, and hundreds of internal Git mirrors, but found it difficult to maintain as the merge rate increased.
Scion assumes that clients are always online. The server is responsible for history, blame, commits, merges, and rebase operations, while the client only maintains a bounded local cache and fetches data on demand. Checkouts are sparsely defaulted, and some jobs can query differences entirely through the API without a full checkout. The working unit is a changeset, stored as a pre-write log entry containing a list of file changes relative to the parent node and content-addressable pointers.
A seven-person team completed the construction and migration of Scion with the help of Claude, while also delivering the code review interface Antique, code search, and source control IAM systems. After migration, the CPU and memory usage of the Scion infrastructure shrank about 200 times compared to the previous Git mirrors, and CI and research job checkout times dropped from minutes to seconds, with CI alone freeing up about 25% of CPU capacity. The company claims that Scion is not a general-purpose Git replacement and is currently limited to internal use.
The daily merge volume increased from 2,500 to over 30,000, making Git based on complete local copies a bottleneck. This event is driven by the scale of internal code generation within the company. Server-side computation and sparse checkouts reduce local and network load, allowing CI and research jobs to achieve faster checkouts and lower resource consumption; external developers and toolchains relying on traditional Git workflows face similar infrastructure pressures under comparable scales.
Source: Public Information
ABAB AI Insight
Anthropic, driven by Claude, has increased its daily merge volume from about 2,500 in May 2026 to over 30,000 by September, with the main branch history nearing 100GB. The company previously managed with partial copies and hundreds of Git mirrors, but the approximately 25 million daily CI checkouts have overwhelmed the traditional architecture. A seven-person team completed the construction and single-repository migration of Scion in five months, while also delivering supporting tools like the Antique review interface.
Resources have shifted from complete client repository copies to a server-dominated sparse and on-demand model. Scion places history, merges, and rebase operations on the server, with the client retaining only a bounded cache and pulling on demand; changesets record incremental changes in a pre-write log, avoiding locks on the entire repository. After migration, the infrastructure's CPU and memory usage shrank about 200 times, checkout times dropped to seconds, and CI freed up about 25% of CPU capacity. The design references Google Piper and Meta Sapling but adjusts for always-online agents and automated jobs.
Similar to Google and Meta's self-built version control systems for ultra-large single repositories, Anthropic replaces Git's offline complete copy assumption with a server-centric architecture. It is currently at a stage where AI agent-generated code scales beyond the design boundaries of human workflows, making source control a major bottleneck following model capabilities.
This is essentially a technological replacement. Git assumes clients are offline and hold complete repository copies, while AI agents and automated jobs are always online and only need partial files. This change occurs because the daily merge volume has grown over tenfold, making the network and computational costs of complete checkouts and mirror synchronization outweigh the benefits. Server-side computation and sparse checkouts shift resource consumption from the client to horizontally scalable cloud infrastructure, allowing high-concurrency submissions to no longer be limited by local repository size and lock contention.
ABAB News · Cognitive Laws
The next bottleneck for agent-generated code is the version control system's copy assumption
Server-dominated sparse checkouts shift local resource consumption to the cloud
The speed of infrastructure replacement completed by a seven-person team with the model exceeds traditional human cycles