GitHub unveiled on November 7 the Rubicon Git storage architecture supporting 5000 concurrent agents in a single repository
Internal stress testing showed the new architecture handling 5000 concurrent agents in one repository and sustaining 7200 pushes per second, now in limited private preview with broader rollout planned for early next year.
On November 7, it was reported that on November 5, GitHub published a blog post saying it had formally announced a brand new Git storage architecture codenamed Rubicon at the GitHub Universe conference the previous week. The company said the architecture not only handles thousands of writes per second but also allows reads to scale on demand. Internal stress testing confirmed that the new architecture can keep reading latency extremely low while supporting 5000 concurrent agents in a single repository and absorbing up to 7200 code pushes per second. The architecture is now in limited private preview and is planned to open up gradually in the early part of next year.
Today the write peak of the busiest GitHub repositories is at most a few dozen per second, and usually appears as a short burst. But as more development work is handed to AI agents running around the clock at high concurrency, those instantaneous bursts are expected to become normalized sustained high-throughput loads within a year.
Every code push cascades into more read requests: continuous integration pipelines start immediately, deployment systems pull the latest commits, and other collaborating agents fetch code as context. This read fan-out significantly amplifies overall platform traffic as write frequency rises, and the trend is already visible at the top of the traffic distribution curve. A typical GitHub repository shown by the company saw daily request volume jump roughly fourfold in the two months between March and May, and access growth has since maintained a steep exponential trajectory.