5,000 agents push code to the same warehouse at the same time, 7,200 times per second, and the read latency is still at a low level. This is the result of GitHub's new Git storage architecture Rubicon in internal testing.
Ordinary teams cannot use this magnitude. But it illustrates how GitHub sees its next users: not dozens of engineers, but hordes of AI agents.
The original design premise of Git was that people were submitting, the frequency was low, and conflicts were resolved by people. A large group of agents modify the same warehouse concurrently, and there will be crazy fights over reference updates, which traditional storage cannot handle.
The disassembly method of Rubicon is as follows:
- Git objects into Azure Blob Storage
- The reference state is managed by Ref-CAS, and concurrent updates are handled using compare and exchange.
- The virtual copy is responsible for reading and writing
- The background node specializes in compressing data packets
It is equivalent to changing Git from "one directory on one machine" to distributed storage on the cloud.
It is currently a limited private preview and will be gradually released. Teams that use GitHub Copilot agent mode, or use tools such as Codex and Cursor to open branches in batches, will most likely benefit indirectly in the future, and push queues and reference conflicts will be much less.