OpenAI's Pando: Rebuilding Git Infrastructure for Agent-Scale Development
OpenAI unveiled Pando, a custom Git backend handling nearly a petabyte of data daily with 5x faster push latency, enabling parallel development at the scale required by AI-assisted coding.
Last verified:
OpenAI’s Custom Git Backend Addresses Agent-Scale Development Demands
OpenAI built Pando, a horizontally scalable Git backend, to handle the explosion in concurrent development work enabled by its AI coding tools. According to OpenAI’s engineering blog, Pando reduced median push latency by nearly 5x and now serves approximately 1 petabyte of repository data per day—a scale that revealed the limits of traditional Git hosting under agent-driven workflows. The system maintains full Git protocol compatibility, allowing both human engineers and AI agents to adopt it without changing their command-line interfaces or development processes.
Why Git Infrastructure Became a Bottleneck
OpenAI’s monorepo infrastructure began failing under pressure around late 2025, with fetch and push operations slowing and reliability dipping below 90%. According to the OpenAI Blog, the root cause was not Git’s design but rather the mismatch between Git’s original use case—kernel patch exchange via email—and the demands of parallel AI-assisted development. Codex and similar code-generation systems enable individual engineers to stage multiple changes for continuous integration simultaneously. When each push triggers dozens of CI jobs fetching the same commit, the system must handle both write concurrency (many branches receiving pushes) and read concurrency (many clients downloading identical history) at once.
Traditional Git hosting solutions optimize for one dimension: GitHub and similar platforms excel at coordinating writes across distributed teams but can bottleneck on reads when a single commit is fetched thousands of times by CI runners and developer machines in parallel.
Pando’s Dual-Axis Design
Pando was engineered with two explicit goals: horizontally scalable writes optimized for many concurrent pushes to different branches, and maximized read throughput to saturate clients’ available network bandwidth. According to OpenAI’s technical writeup, the system decouples the write path (branch updates) from the read path (history downloads), allowing each to scale independently. This separation lets OpenAI add write capacity by distributing branch state across servers and read capacity by distributing history chunks across a content-delivery layer.
The constraint OpenAI imposed on itself—maintaining Git protocol compatibility—was deliberate. Rather than replacing Git with a proprietary system, the team rebuilt the backend while keeping the interface identical, so existing client tools, CI pipelines, and emerging AI agents could adopt Pando without modification. This choice also enabled OpenAI to contribute performance improvements back to the upstream Git project, extending the benefit beyond their organization.
Why This Matters
The infrastructure pattern Pando establishes has implications beyond OpenAI’s walls. As AI-assisted development tools mature, teams will face similar scaling problems: AI agents and humans generating code in parallel, multiple CI systems fetching the same commits, and development velocity that outpaces traditional hosting capacity. OpenAI’s decision to optimize for this specific load profile—not just aggregate throughput, but write concurrency and read saturation—signals that agent-scale development infrastructure will differ materially from the monorepo tooling built for human-scale teams.
The durability of Pando’s Git compatibility matters for adoption risk. Because it speaks the Git protocol, teams could experiment with Pando or switch to it without retraining workflows or rewriting scripts—a critical factor if AI-native development tooling is to achieve mainstream adoption. OpenAI’s stated intent to upstream performance patches also suggests a path where improvements to Git infrastructure are shared across the ecosystem rather than locked into proprietary platforms.
Frequently Asked Questions
Why did OpenAI build a new Git system instead of using existing solutions?
OpenAI needed to scale Git for agent-driven development where many branches receive pushes in parallel and many CI systems fetch the same commit concurrently. Existing Git hosting solutions struggled to meet these dual requirements at OpenAI's scale.
Do developers need to change how they use Git with Pando?
No. Pando retains full Git protocol compatibility, so engineers and AI agents can adopt it seamlessly into existing workflows without learning new commands or interfaces.
What was the performance improvement?
Pando reduced median push latency by nearly 5x and can deliver up to 1 petabyte of Git data per day while maintaining reliability above 90%.
Will Pando be open-sourced?
OpenAI plans to contribute improvements back to upstream Git, though the blog does not announce a public release timeline for Pando itself.