When the bottleneck stopped being the humans
Anthropic built Scion because its automated development workload at scale made Git’s distributed repository model an expensive bottleneck, not because Git became poor software. The challenge wasn't Git's core functionality, but its architectural cost when confronted with an entirely new class of user: AI.
Claude’s relentless pace pushed Anthropic’s main-branch history to nearly 100 GB. Its CI infrastructure was performing roughly 25 million checkouts per day. This unprecedented scale exposed a critical weakness in traditional version control for AI-driven development.
Human developers typically clone repositories once, then work locally for extended periods. But always-connected agents and ephemeral CI jobs repeat repository work with extreme frequency. This shifts the constraint from human typing speed to the underlying infrastructure, as each clone and local operation consumes significant compute and bandwidth. Anthropic needed a system built for machines, by machines, that could outpace even Claude.
Git’s strength became its scaling tax
Git’s distributed architecture, while a boon for human collaboration, became a scaling tax for Anthropic’s automated workload. Each client typically maintains a complete local copy of the repository’s history, enabling offline work and rapid local operations. This design excels for human developers, who benefit from full context and infrequent, complex changes.
However, Claude’s CI infrastructure performed approximately 25 million checkouts daily. Most of these clients were ephemeral, always-connected agents and automated jobs. These workers rarely needed the full 100-gigabyte main branch history, instead requiring only selected files for brief, repetitive tasks. Fetching and maintaining redundant local copies for each short-lived process became a massive overhead.
This pattern of high-volume, low-context operations directly contrasted with Git’s strengths. The system’s design, optimized for human teams where local history and offline work are invaluable, proved inefficient for automated workers that needed only specific data. Git was not inherently broken; its architecture simply no longer aligned with Anthropic’s extreme, automated scale.
Scion moves the heavy lifting server-side
Scion, Anthropic’s in-house version control system, represents a fundamental shift in how large-scale codebases are managed. Instead of maintaining large local repositories, Scion adopts a thin-client approach: clients request only the files and data they need, while commands, merges, and history operations are offloaded to powerful servers. This centralization drastically reduces the load on individual client machines.
This server-side heavy lifting yielded impressive results for Anthropic. According to their reports, Scion uses around 200 times less CPU and memory than their previous Git mirroring infrastructure. Critical checkouts, once taking minutes, were reduced to mere seconds, and the company reclaimed approximately 25% of its CI CPU capacity.
It is crucial to understand the context of these claims. These are workload-specific metrics from an environment where Claude-driven programming loops generated over 30,000 merges per day and CI performed 25 million checkouts daily. This scale is beyond what most organizations encounter.
These figures do not imply that Scion makes every Git task 200 times faster for typical development teams. Git remains an excellent solution for the vast majority of users, and Anthropic’s innovations address a niche problem arising from AI-driven hyper-scale. For a deeper dive into their architectural choices, consult their engineering post: Redesigning source control at Anthropic.
Нравится статья? Получайте такие каждое утро на почту.
одно письмо в день · отписка в два клика · без сторонних трекеров
The real price—and the lesson for Git
Replacing Git meant more than just a new version control system; it demanded a complete re-engineering of Anthropic's developer toolchain. The team rebuilt critical infrastructure, including: - Antique, their custom code review interface - Bespoke code search and repository-level access controls - New automated landing and merge queue systems
This was an infrastructure bet for an exceptional monorepo and agent workload, not a general recommendation. Anthropic’s CI infrastructure performed 25 million checkouts daily, with a main branch history approaching 100 gigabytes. Scion’s 200x reduction in CPU and memory usage, and checkouts shrinking from minutes to seconds, directly addressed their unique scale.
AI-driven code volume places unprecedented pressure on existing checkout, testing, and review systems. While Scion delivered a 25% recovery of CI CPU capacity for Anthropic, most teams will not encounter this extreme bottleneck. Instead, Git's future will likely involve enhanced server-side capabilities and more efficient partial-transfer protocols, rather than universal replacement.
Frequently Asked Questions
What is Scion?
Scion is Anthropic’s server-centric version control system, designed to serve files and perform repository operations without requiring every client to maintain a full local repository.
Why did Anthropic build Scion instead of scaling Git?
Anthropic’s high volume of automated jobs and CI checkouts made Git mirroring costly. Scion was designed to reduce duplicated data and work across those clients.
Is Scion 200 times faster than Git?
No. Anthropic reported around 200 times less CPU and memory use than its previous Git mirroring infrastructure; that is not a general speed comparison with Git.
Should most teams switch from Git to Scion?
No. Scion targets Anthropic’s unusually large, automation-heavy workload. Git remains a strong fit for most development teams.