# GitHub Tests Rubicon With 5,000 AI Agents at 7,200 Pushes Per Second

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/52823
Published: 2026-11-10T05:24:47.000Z
Updated: 2026-11-10T05:24:47.000Z
Section: Technology

> The test ran thousands of agents in one repository while keeping read latency low. Rubicon remains in limited private testing, with broader rollout planned for early 2027.

GitHub tested Rubicon, a storage architecture for AI coding workloads, at 7,200 code pushes per second in one repository during a test with thousands of agents.

The run involved 5,000 AI agents working in that repository and maintained low read latency. Rubicon remains in limited private testing, with a broader rollout planned for early 2027.

A Git push handles two tasks: saving new or changed code and moving a branch pointer to its latest version. Git stores changed content as new objects instead of overwriting earlier versions, and unchanged parts can be reused. Rubicon's design addresses both the work of updating branches and the work of checking for changes.

Agents working on separate branches can update those branches independently. If two agents try to update the same branch, the system still needs to resolve the conflict.

GitHub's earlier design made every branch use one shared update record. That meant separate branches could compete for the same record, limiting throughput to hundreds of pushes per second. A later design split the records to allow more concurrent updates, but clients then had to check each record to find out whether the repository had changed, slowing reads.

Rubicon uses Ref-CAS to coordinate branch updates. It allows different branches to update concurrently while still handling conflicts on a shared branch. The system also assigns the whole repository a version number, so a client can check for changes without querying every branch record.

Separate machines merge accumulating data packs in the background to keep them from slowing reads. The version check and background merging address the read side of the workload as agents add code and branch updates build up.

AI coding agents can generate a steady stream of commits. Each commit may trigger automated tests and allow other agents to download the latest version, adding read activity alongside writes. GitHub's busiest repositories previously saw only brief bursts of dozens of writes per second.