GitHub Announces Rubicon Architecture; Internal Tests Show Support for 5,000 Concurrent AI Agents
GitHub released Rubicon, a new Git storage architecture designed to handle high-concurrency AI agent writes, with internal tests showing support for 7,200 pushes per second.
On November 9, 2026, GitHub announced its next-generation Git storage architecture, "Rubicon," to address the demand for large numbers of AI agents operating on a single code repository simultaneously. According to iThome, official internal tests show the new architecture can support 5,000 AI agents working concurrently, processing 7,200 code pushes per second while maintaining low read latency. The new architecture is currently available only in limited private preview.
As AI-assisted and agentic development have become major factors in rapid platform traffic growth, GitHub has previously faced infrastructure load issues. In August, GitHub experienced a nearly 8-hour service outage due to record-high traffic and key infrastructure failing to scale in time. GitHub stated that currently, the busiest repositories see at most dozens of writes per second during short peaks, but expects these brief write spikes to become sustained loads over the next year.
The Rubicon architecture separates data storage, read/write computation, and reference updates. GitHub uses its existing request routing and load-balancing service, Spokes, to distribute read/write requests to virtual replicas. Git objects are stored as packfiles in Azure Blob Storage, while branch reference information is managed by Ref-CAS, a reference management service developed by GitHub. Ref-CAS manages reference updates via a Compare-and-Swap mechanism, allowing different AI agents to update their respective branches in parallel without waiting for sequential writes across the entire repository.