GitHub Unveils Rubicon Git Architecture to Support High-Concurrency Writes from AI Agents

To address infrastructure loads driven by AI-assisted and agentic development, GitHub announced Rubicon, a next-generation Git storage architecture. Internal benchmarks indicate that Rubicon supports 5,000 AI agents operating simultaneously on a single repository, handling 7,200 code pushes per second while maintaining low read latency. The architecture is currently in limited private preview. GitHub reported that monthly code commits grew from 1.4 billion in April to 2.9 billion, with high-traffic repository requests surging significantly. The platform anticipates that short-lived write spikes will transform into continuous workloads as AI agents operate around the clock. In technical design, Rubicon decouples storage, compute, and reference management: Virtual Replicas process repository read/write operations while holding only caches; Git objects are stored as Packfiles in Azure Blob Storage; the proprietary Ref-CAS service manages branch and reference updates via Compare-and-Swap, locking only targeted references so agents can update distinct branches in parallel; pack consolidation and garbage collection are offloaded to independent nodes to prevent maintenance tasks from degrading real-time performance.

Key takeaways

• GitHub internal tests show Rubicon supports 5,000 concurrent AI agents and handles 7,200 code pushes per second.
• The architecture decouples storage, compute, and references, storing Git objects in Azure Blob Storage while replicas maintain only caches.
• The proprietary Ref-CAS service uses a Compare-and-Swap mechanism, enabling multiple AI agents to update distinct branches in parallel.
• Maintenance tasks such as pack consolidation and garbage collection are offloaded to dedicated nodes to ensure stable latency.