What changes compared with today
GitHub presented a new Git storage architecture codenamed Rubicon at GitHub Universe, as reported by IT之家. According to the company, the system can handle thousands of writes per second and scale read capacity horizontally on demand while keeping latency low.
The comparison offered by the source is with the current state: the platform's busiest repositories see write peaks on the order of a few dozen per second, and these are generally short bursts. Rubicon is designed for a different, continuous, and much more intense load profile.
The internal test numbers
In stress tests conducted internally, the architecture sustained the simultaneous collaboration of 5,000 agents on a single repository, with a peak of 7,200 pushes per second. These are results stated by GitHub and not independently verified.
The company also claims that the new structure allows it to maintain very low read latency even under these conditions. Availability is currently limited to a restricted private preview: the company plans to progressively broaden access starting at the beginning of next year.
Why the load is growing
GitHub's explanation links the need for Rubicon to the adoption of AI agents in software development. According to the same source, within a year occasional peaks should turn into a high, continuous load, because agents operate constantly and concurrently.
An amplifying factor is the cascading effect of reads: each push triggers continuous integration pipelines, causes deployment systems to pull commits, and prompts other agents to synchronize the code as working context. The increase in writes therefore translates into a much larger number of read requests.
GitHub notes that the phenomenon is already visible in the most heavily used repositories: one cited case shows daily requests quadrupling in about two months, between March and May, with growth that the source describes as still steep.
What remains to be verified
The performance data come exclusively from internal tests and not from external measurements. It does not state which repositories or configurations were used, nor how the 5,000 agents were simulated. It also remains unclear when and how the preview will be expanded and what impact it will have on commercial plans, aspects the material does not address.
