Anthropic has replaced Git for its main shared codebase with an internally developed system called Scion after its daily code change volume surged from roughly 2,500 in May to more than 30,000 by September. The rapid expansion of autonomous AI coding agents powered by Claude accelerated development to the point where managing, reviewing, and testing changes eclipsed code authoring as the primary bottleneck. A team of seven engineers developed Scion, alongside a companion code review system called Antique, over a five-month span using Claude to write substantial portions of the codebase under human supervisory control.
Shifting the Bottleneck from Code Generation to Infrastructure
The transition occurred as the operations required to support each software submission began straining the company’s internal developer infrastructure. Automated testing systems at the company execute approximately 25 million code retrievals each day to validate incoming modifications. Operating hundreds of local mirrors of the Git repository had become increasingly fragile and resource-intensive to synchronize at that rate of modification.
Scion mitigates the operational overhead by moving workload execution onto centralized servers, permitting automated pipelines and engineers to query only targeted files or metadata. Automated tools can evaluate history or generate diffs without pulling full local repository clones. Following the deployment, Anthropic reported that code retrieval operations decreased from minutes to seconds, cutting required CPU and memory consumption to one two-hundredth of the footprint used by its former Git mirroring infrastructure.
Human Oversight and Architectural Constraints
While Claude generated significant volumes of implementation code, human engineers directed architectural boundaries, system integration, and edge-case management. Anthropic Chief Technology Officer Rahul Patil noted on LinkedIn that the team was responsible for establishing core design trade-offs, which frequently required overriding Claude when the model suggested approaches violating critical design rules. In one instance, generated code attempted to scan entire version histories to retrieve data that was already directly accessible through the review platform.
Engineering attention has subsequently shifted from granular line-by-line inspection toward evaluating broader intent, security boundaries, and architectural coherence. To maintain stability amidst rapid commit velocity, the team built isolated simulation environments, fuzz testing pipelines, and rate-limiting safeguards. The system also introduces fine-grained directory permissions and review tooling that distinguishes cosmetic modifications, such as comment edits, from underlying business logic.
Bespoke Scale Rather than a Universal Replacement
Anthropic underscored that Scion was engineered specifically to handle its own high-throughput environment, which involves massive agentic workloads and an unusually large shared repository. Adopting the platform outside Anthropic would demand wholesale structural overhauls across existing search, code review, and automated testing stacks.
The company clarified that Git remains the optimal version control platform for typical software teams, confirming that Scion is neither intended as a general replacement for Git nor structured as an off-the-shelf public developer product. Rather than distributing the platform as open-source tooling, Anthropic is sharing design findings and architectural methodologies that standard Git deployments can adapt to streamline central processing and curb redundant data transfers.
Follow for daily updates on cybercrime, corporate fraud, DFIR, hacking, investigations, and digital forensics
