Anthropic Replaces Git with 'Scion' Amid AI Code Boom, Builds Custom 'Antique' Review System
As AI coding tools drove daily code changes from 2,500 to over 30,000, Anthropic deployed a 7-person team over 5 months to migrate from Git to Scion and built a
As internal adoption of AI coding tools triggered a dramatic surge in code generation, Anthropic fundamentally overhauled its version control architecture, moving from Git to an internal source control system named 'Scion' and building a dedicated code review platform called 'Antique'. On October 13, 2026, independent technologist Simon Willison analyzed Anthropic's transition, detailing how a seven-person engineering squad resolved acute repository bottlenecks through a focused five-month infrastructure migration.
Image source: @kaigooddev via X / Simon Willison
This infrastructure shift serves as a concrete case study of how software development at frontier AI laboratories is confronting the architectural boundaries of Git and conventional pull request (PR) workflows under extreme, machine-driven code production rates.
30,000 Daily Code Changes: The Limits of Git Infrastructure in the AI Era
The immediate catalyst for Anthropic's architectural overhaul was an exponential spike in code modification volume.
With internal adoption of AI coding tools accelerating rapidly, daily code changes surged from an average of roughly 2,500 to over 30,000 per day. This dramatic volume leap severely tested conventional version control systems originally optimized for human-scale iteration tempos.
- VCS Throughput Bottlenecks: Funneling tens of thousands of daily code changes into a single monorepo triggered severe throughput and synchronization latency within standard Git workflows.
- Review Pipeline Overload: Because code generation velocity drastically outstripped human review capacity, the standard PR-based review pipeline faced acute strain. Line-by-line manual inspections in traditional review UIs struggled to absorb massive machine-authored diffs.
A 5-Month, 7-Person Migration: Moving to Scion and Building 'Antique'
To untangle these operational logjams, Anthropic assembled a dedicated team of seven engineers, completing a comprehensive migration of their core internal codebase from Git to 'Scion' over a five-month span.
Concurrently, to address the mounting backlog and strain associated with reviewing massive volumes of AI-generated code, the team developed and rolled out 'Antique', a custom in-house code review platform.
- Adopting Scion: By deploying a source control engine suited to handle massive monorepos and ultra-high-frequency concurrent write traffic, the team resolved centralized synchronization and repository bottlenecks.
- Deploying Antique: Built specifically to accommodate the velocity of machine-generated code, Antique provides a dedicated review platform. Recognizing that changing tools does not automatically improve review quality, the initiative focuses on rethinking the review process itself.
Practical Boundaries and Takeaways for Engineering Organizations
While Anthropic's transition illustrates proactive infrastructure adaptation in the AI era, several practical boundaries distinguish this specialized case from standard engineering practices:
- Proprietary Internal Tooling: Both Scion and Antique were engineered specifically for Anthropic's internal monorepo topology and operational infrastructure. Neither tool is currently available as a turnkey, public open-source release for external engineering teams.
- Tooling Does Not Guarantee Review Rigor: Replacing the underlying VCS engine with Scion resolves infrastructural throughput limits, but it does not inherently guarantee the semantic correctness or reliability of AI-generated code. Substantive review process redesign remains mandatory alongside tooling changes.
- Scale-Dependent Requirements: Organizations operating below extreme thresholds—without tens of thousands of daily code changes—will find standard Git ecosystems, GitHub/GitLab PR flows, and conventional CI/CD pipelines more than sufficient for everyday software delivery.
Sources
- Simon Willison's Weblog: Redesigning Source Control at Anthropic
- Kai Good (@kaigooddev) on X: 2026-10-15 Commentary