OpenAI has also taken action to transform Git: Codex participated in rebuilding the backend, launched in 10 weeks, with pushes nearly 5 times faster.

Beating AI News Flash: AI is writing code faster and faster, but Git backends are starting to buckle under the pressure. Following Cursor and Anthropic, OpenAI has also unveiled its in-house Git backend Pando, which has now taken over the company's entire main code repository.

Over the past year, Codex has enabled engineers to dispatch large numbers of Agents to write code simultaneously, causing code branches, pushes, merges, and pull requests to surge. OpenAI's original system once frequently stalled, merges failed, and reliability even dropped below 90%. So one engineer took charge of core architecture design, multiple Codex Agent teams participated in implementation, and Pando was put into production in just 10 weeks.

Pando allows code pushes from different branches to be processed in parallel without queuing behind each other. When pulling code, it no longer generates separate data packets for each client, but reuses existing Git data packets and distributes them directly through caching. After the overhaul, median code push latency was cut by nearly 80%, nearly 1 PB of data is distributed daily, and the cache hit rate exceeds 99%. However, different branches are no longer guaranteed to present the state at the same moment.

The three companies are solving similar scaling challenges, but their choices are completely different: Cursor's Origin is still based on Git and mainly aims to replace GitHub's code hosting service; Anthropic's Scion completely abandons Git, moving even commits, merges, and history queries to the server side. OpenAI chose to keep the Git client and protocol and only rebuild the backend, so engineers and Agents basically do not need to change their original development workflows.