AI News
All stories
Reported

OpenAI built an internal code repository to replace Git for agent-driven commits

OpenAI's CTO of Applications, V. J. Raji, says AI agents now push code faster than Git was designed to handle, so OpenAI built its own system.

AI in Practice · · 2 min read

OpenAI has built its own internal code repository system. According to V. J. Raji, its CTO of Applications, AI agents now push code faster than Git was designed to handle. In an interview that Nate B Jones shared, Raji dates the change to February 2026, when powerful models, Codex, and many agents working in parallel on separate work trees came together. He says the number of merge requests engineers created "simply skyrocketed" after that point.

Raji offered token usage as a rough gauge of how deeply AI is used. He said it is not necessarily an accurate measure of engineer productivity, but that it shows how broad and deep AI usage is in a given area. At frontier labs, he said, token usage rose 17 times over the past year, compared with about 2 times at regular software companies.

In Raji's account, the bottleneck has moved from writing code to what comes after. He named code review, testing, and CI/CD as the places where work now piles up. OpenAI is addressing them with an auto-check system that teams can enable. He also said more code is now thrown away, since prototypes are cheap to build and easy to reject.

Raji was most specific about the repository problem. He said OpenAI uses a monorepo, and that the volume of code passing through it is already an order of magnitude greater than the system was designed for. Git assumes offline work, he explained: a developer downloads the code, works on local branches, and commits when finished. Agents are always online, so OpenAI's internal team built a system around that assumption. Raji said the new system supports the monorepo and that all engineers now use it. He did not name the system or describe its design beyond that.

Raji also described two further problems. Keeping the main branch stable has traditionally meant reconciling changes before they land, which becomes hard when many agents commit constantly. He said locking the master branch on every transaction would undercut throughput, so conflicts have to be handled differently. He also called the speed at which agents load code a serious problem, and said the fix starts with sparse checkout, along with caches so agents can download data without constantly hitting the main server. The interview doesn't say how conflicts are resolved in practice, or how much traffic the caches remove.

What changed

OpenAI built its own internal code repository system to replace Git for agent-driven commits.

Type: Other · Applies to: monorepo

Why it matters

Companies running many coding agents may find that version control, code review, CI/CD, and testing become bottlenecks before code generation does. Raji's account suggests these limits show up quickly once agents commit continuously.

What you can do

Measure where review, testing and CI/CD queues slow down as agent-generated commits increase, before adding more agents.

Original source

We have no link to the original source for this story yet.

Where we saw it

1 source

Nate B Jones

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works