OpenAI is also revamping Git: Codex is contributing to the backend overhaul, with the updated system set to launch in 10 weeks and deliver code pushes nearly five times faster.

Dongcha Beating AI News Flash: As AI coding accelerates, Git backends are hitting capacity limits. Following Cursor and Anthropic, OpenAI has unveiled its self-developed Git backend Pando, which now manages the company’s entire main code repository. Over the past year, Codex has enabled engineers to deploy numerous AI agents to write code simultaneously, leading to a surge in code branches, pushes, merges, and pull requests. OpenAI’s legacy system experienced frequent lags and merge failures, with reliability dropping below 90% at one point. Subsequently, one engineer led the core architecture design, with multiple Codex Agent teams contributing to implementation, and Pando was launched into production in just 10 weeks. Pando enables parallel processing of code pushes across different branches, eliminating the need for queuing. When pulling code, it no longer generates individual data packets for each client; instead, it reuses existing Git data packets and distributes them directly via caching. After the upgrade, the median latency for code pushes was reduced by nearly 80%, with nearly 1 PB of data distributed daily and a cache hit rate exceeding 99%. However, the system no longer guarantees that different branches reflect the same state at a given time. The three companies are tackling similar scaling challenges but have taken vastly different approaches: Cursor’s Origin remains Git-based, primarily aiming to replace GitHub’s code hosting service; Anthropic’s Scion abandons Git entirely, moving commits, merges, and history queries to the server side. OpenAI opted to retain the Git client and protocol, rebuilding only the backend, so engineers and agents can largely keep their existing development workflows unchanged.