OpenAI builds custom Git backend in just 10 weeks to handle 2,442% code commit surge

OpenAI has developed a new Git backend called Pando to handle the growing volume of code changes generated by its engineers and Codex agents. The system now powers development across OpenAI’s monorepo, processing nearly one petabyte of Git data daily. The infrastructure addresses a problem that emerged as Codex accelerated software development. More engineers and autonomous agents now work on separate code changes simultaneously. Their activity creates a growing stream of branch updates, code reviews, merges, and repository downloads. OpenAI says Pando has reduced median Git push latency by nearly 80 percent. It has also improved merge reliability to 99.99 percent, compared with earlier periods when reliability sometimes dropped below 90 percent. Codex pushes Git to its limits OpenAI’s monorepo experienced a sharp increase in development activity over the past year. Its quarterly commit growth reached 2,442 percent in Q4 2026 compared with Q1 2024. Traditional Git infrastructure struggled to accommodate the resulting workload. Engineers encountered slower pushes and downloads, alongside failed merges. Multiple automated jobs also fetched identical commits after individual code submissions. We built Pando, a new Git backend for our monorepo, to handle more pushes and fetches without changing how our engineers and agents use Git.Pando cut push latency by ~80% and now serves nearly a petabyte of Git data a day. We can grow Git capacity alongside agent usage. pic.twitter.com/4c0YXOfKqK— OpenAI Developers (@OpenAIDevs) November 3, 2026 OpenAI wanted to address these bottlenecks without replacing Git or forcing developers to adopt unfamiliar tools. Its engineers therefore designed Pando around the existing Git protocol and command-line client. The approach preserves familiar development workflows while changing how servers coordinate updates and distribute repository data. It also allows improvements to the underlying Git client to benefit the wider developer community. Conventional Git infrastructure can struggle when thousands of developers update different branches simultaneously. Some distributed systems coordinate changes across multiple repository replicas. Others rely on a shared write-ahead log to establish update order. Both approaches introduce potential bottlenecks. Coordinating updates across replicas increases communication overhead. A shared log can also force unrelated branch updates to compete for the same resource. Pando takes a different approach through causal consistency. It coordinates updates independently for individual branches instead of imposing one global sequence. When a developer pushes code, Pando stores the incoming Git packfile in cloud object storage. Its metadata service then records the updated branch version and its dependencies. This arrangement allows unrelated branches to accept updates concurrently. However, readers examining multiple branches might see different update versions. Individual branch updates remain consistent. Pando also reuses previously generated packfiles during downloads. This reduces the computing work required to serve identical repository data repeatedly. Caches accelerate repository downloads The architecture uses regional caching to distribute Git data closer to developers and automated systems. Its Express delivery service stores packfiles in regional caches. OpenAI reports cache hit rates exceeding 99 percent. The company has also demonstrated clone and fetch operations saturating 10-gigabit-per-second network connections. Background compaction further improves performance. Workers combine smaller packfiles into larger groups, reducing download overhead and speeding up repository history searches. Built with Codex in 10 weeks OpenAI says it developed Pando with Codex in just 10 weeks. A single engineer designed its core architecture, while teams of agents handled implementation. The system entered production in April 2026. OpenAI implemented it in Rust and developed extensive testing tools to identify concurrency bugs. Pando has since handled more than 57 terabytes of Git packfiles in a single hour. The company says it achieved this using more than 99 percent locally cached data. OpenAI also plans to explore broader applications for Pando. Its architecture could support other environments where large numbers of software agents collaborate on shared repositories.Get the latest in engineering, tech, space & science - delivered daily to your inbox.Aamir is a seasoned tech journalist with experience at Exhibit Magazine, Republic World, and PR Newswire. With a deep love for all things tech and science, he has spent years decoding the latest innovations and exploring how they shape industries, lifestyles, and the future of humanity.