[GitHub Rebuilds Git's Underlying Structure: 5000 AI Agents Work in Parallel, Handling 7200 Code Pushes Per Second in a Single Repository] Beating AI News: GitHub has announced a new generation of Git storage architecture, codenamed Rubicon. Internal tests show that 5000 AI agents can work simultaneously in a single code repository, completing 7200 code pushes per second while maintaining low read latency. Previously, even GitHub's busiest code repositories only saw a few dozen writes per second for short periods. Now, a large number of AI agents can develop in parallel around the clock, constantly committing new code. Each commit may also trigger automated tests, causing other agents to download the latest version, increasing the read and write pressure on the repository. GitHub's solution is to separate "storing code" and "updating branches." Each time Git pushes code, it performs two tasks. The first step is to save the added or modified code content. Git does not directly overwrite the old version; instead, it stores the changed content as a new object, allowing the unchanged parts to be reused. The second step is to update the branch pointer, recording which latest version the branch now points to. For example, 1000 agents might be developing software simultaneously, some modifying login functionality, others payment functionality. As long as they work on different branches, new code can be saved simultaneously, and each branch can be updated independently. The system only needs to determine which update succeeds first when two agents are simultaneously vying to update the same branch. Initially, GitHub shared a single update record across all branches. Even when different AI agents modified different branches, they still competed to update the same file, resulting in a processing limit of only a few hundred pushes per second. GitHub then tried splitting the record into multiple parts, allowing different branches to update simultaneously. However, this required querying each branch individually to check for changes, slowing down read speeds. Ultimately, GitHub developed the Ref-CAS service specifically for managing branch updates. It allows simultaneous updates across different branches, handling conflicts only when modifying the same branch. It also maintains a unified version number for the entire repository. Other programs only need to check this version number to know if the repository has changed, without having to query each branch individually. GitHub deployed a separate machine to merge the accumulating code data packages in the background, preventing the data packages from growing too large and slowing down read operations. The new architecture ultimately achieved 7200 pushes per second in internal testing. GitHub's Rubicon is currently in limited private testing and is planned for widespread deployment in early 2027. [Original Link]