Why OpenAI Rebuilt Git From Scratch (And What It Means for AI-Assisted Building)

OpenAI built and deployed a custom Git backend called Pando to stop their internal code repository from collapsing under the weight of Codex. The short version: AI agents were pushing so many simultaneous code changes that standard Git infrastructure could not keep up, so the team rebuilt the plumbing from scratch.

That sounds like a deeply internal engineering problem. It is not. It is a signal worth paying attention to if you are building anything with AI today.

The Problem Nobody Talks About: AI Moves Fast, Infrastructure Does Not

Git is the near-universal system developers use to track and save code changes over time. OpenAI's engineers work inside a single large shared repository, what engineers call a monorepo, where every team's code lives together. As Codex began running autonomously on real tasks, the volume of concurrent pushes, merges, and fetches ballooned. By late last year, reliability had dropped below 90% at times, meaning roughly one in ten operations was failing. Slow fetches, failed merges, and push timeouts became routine friction.

The root cause was architectural. Traditional Git assumes a relatively small number of humans making changes in a loosely coordinated way. AI agents do not work that way. They generate many branches and many pushes at once, often in parallel, which exposes limits that human-paced development never hit.

So OpenAI built Pando, a new backend that speaks the standard Git protocol (meaning engineers and agents do not change a single habit) but handles writes to different branches independently and in parallel rather than funnelling everything through one queue. The results were significant: median push latency dropped by nearly five times, and the system now delivers close to a petabyte of Git data every day.

What This Actually Tells Designers Who Are Starting to Build

Here is the thesis: AI coding tools like Codex are not just faster autocomplete. They are parallel workers that multiply the pace and volume of changes in a codebase. The tooling and infrastructure around them has to evolve to match.

For designers exploring vibe-coded projects or building their own apps with AI, this matters in a practical way. When you use an AI coding agent, you are not making one change at a time and carefully reviewing it. You are often running several threads of work simultaneously, or iterating rapidly across different features. That pattern puts pressure on every layer of your workflow, not just the AI model itself.

The fact that OpenAI needed to rebuild foundational infrastructure to support this pace is a sign that the tools and platforms you use will be going through similar changes. GitHub, GitLab, and other hosting services will face the same pressure as AI-assisted development becomes standard.

What to Watch For in Your Own Workflow

You probably do not run a monorepo with petabyte-scale traffic. But the patterns here translate directly to smaller-scale building.

Branch sprawl is real. When an AI agent works on your project, it often creates branches for every experiment or fix. If you are using GitHub or a similar service, start paying attention to how many open branches you accumulate. Clean them up regularly. Dead branches create confusion for both you and any AI tool that needs to read the state of your repo.

Parallel changes need clear separation. Pando's key design choice was letting updates to unrelated branches proceed independently. You can apply the same logic manually: keep distinct features or fixes on distinct branches, and avoid letting an AI agent touch multiple unrelated things in a single commit. It makes reviewing and rolling back much safer.

Your Git host matters more than it used to. OpenAI's team noted that reliability fell below 90% before they intervened. If you are relying on AI agents to push changes automatically, a flaky Git host becomes a real blocker, not just an inconvenience. It is worth checking whether your current setup has any rate limits or concurrency constraints that could cause silent failures.

OpenAI is contributing changes back to open-source Git. They maintain a public fork and intend to upstream improvements. That means the broader Git ecosystem should get faster and more agent-friendly over time, which is good news for everyone building with AI tools.

A Grounded Takeaway

Pando is not a product you can go and use today. It is internal infrastructure, and OpenAI has not announced plans to offer it externally. So the immediate practical impact for most builders is indirect.

What the story confirms, though, is that the pace of AI-assisted development is genuinely novel, novel enough that even OpenAI had to rebuild foundational tools to keep up. The companies and platforms building the infrastructure layer for AI coding are going to keep evolving fast, and some of those changes will ripple into the tools you already use.

For now, the practical move is simple: treat your repository hygiene as seriously as you treat your design files. Clear structure, disciplined branching, and a reliable host are not just developer habits. They are the foundation that lets AI work quickly without creating a mess you cannot untangle later.