Anthropic rebuilds its coding infrastructure as daily changes pass 30,000

Seven engineers developed the replacement in five months with Claude’s help, while retaining responsibility for design, testing and knowing when to reject its suggestions

Anthropic has replaced Git for its main shared codebase with Scion as daily code changes exceed 30,000. Illustrative image.

Anthropic says Claude helped it produce code so quickly that managing the changes became a bigger obstacle than writing them. The company has now replaced the system behind its main shared codebase after daily changes rose from around 2,500 in May to more than 30,000 by September.

The replacement, Scion, tracks changes to software and brings contributions together so they can be reviewed, tested and used. It takes over from Git, which Anthropic says remains the right choice for most teams but had become difficult to operate at its scale.

Seven engineers built Scion, a companion code review tool called Antique and several related systems, completing the migration in five months. Claude wrote much of the code, but the team remained responsible for deciding how the systems should work together.

“The engineers owned the design and the tradeoffs, and a lot of the job was telling Claude no when it reached for a pattern that broke a core constraint,” Rahul Patil, Anthropic’s Chief Technology Officer, wrote on LinkedIn.

Writing code faster creates a different problem

Every software change needs supporting work: retrieving the right files, checking what has changed, running tests and combining it with other contributions. Anthropic says those processes became its main development bottleneck as its use of AI coding agents grew.

Its automated testing systems alone retrieve code around 25 million times a day. The company had been running several hundred internal copies of its Git repository to meet demand, but keeping them updated and working reliably became increasingly burdensome.

Scion changes how that information is handled. More of the work happens on central servers, and automated tools request only the files or information they need. Some can check a change’s history or compare versions without downloading a local copy of the code at all.

Anthropic reports that code retrieval for testing and research jobs now takes seconds rather than minutes. It also says the replacement uses one two-hundredth of the CPU and memory required by its previous Git mirroring infrastructure.

These are results from Anthropic’s own deployment, where large numbers of AI agents and automated jobs work on an unusually large shared codebase.

Knowing when to tell Claude no

The company’s account also describes what remained difficult even when generating code became faster.

During development, Claude sometimes chose an approach that could complete an individual task but conflicted with the wider system’s design. In one example, generated code searched through the history of changes to find information that was already available from the review tool. Engineers had to redirect it toward the more appropriate method.

AI-assisted reviews could also introduce unnecessary complexity. Anthropic describes a case where concerns about trusting another internal service led Claude to add excessive checks.

The engineers continue to review every change request for Scion. According to the company, their attention has shifted toward understanding the purpose of a change rather than reading every line of code.

Quality controls were developed alongside the system. The team created isolated simulations of infrastructure interactions, tests designed to expose faults using varied inputs, and limits on request rates to help protect the service from overload.

Scion also allows access restrictions on individual directories, protecting sensitive code within the shared repository. Its review tools distinguish edits that only change comments from changes to sensitive program logic, helping reviewers focus their attention.

The process still has tradeoffs. With so many changes arriving while tests are running, Anthropic has invested in further automated checks after changes are combined and before software is deployed. It acknowledges that simultaneous changes can conflict.

A specialized system, not a replacement for everyone

Anthropic is clear that its experience should not be read as a reason for most teams to abandon Git.

Scion is designed for enormous codebases used by tens of thousands of engineers or large numbers of coding agents. The company also acknowledges that it can spend far more on internal AI use than most businesses would realistically choose to spend.

“Scion is by no means a suitable replacement for git for general use,” its engineering account states.

Moving another organization onto the system would require changes across code review, search and automated testing. Anthropic’s tools are highly customized to its own operations.

The company is sharing the design and lessons from the migration, including ways existing Git services could handle more work centrally and reduce unnecessary downloads, rather than presenting Scion as a ready-to-adopt public product.