OpenAI has built Pando, a new Git backend, with Codex agents in just 10 weeks, replacing the system behind its main shared code repository without stopping engineering work. The system moved into production in April 2026 and now supports all development in OpenAI’s monorepo.
Pando Reworks How OpenAI Moves Git Data
OpenAI says the new backend keeps the Git protocol and familiar workflows while changing how servers store and deliver data. One engineer designed Pando’s core, teams of agents implemented it, and a small group of engineers deployed it. The company says agents now run around the clock to red-team both Git and Pando, testing the software that supports its growing AI-driven development workload.
Codex increased the number of branches, pushes, merges, and downloads hitting OpenAI’s source control systems. One engineer could prepare several changes at once, and each push triggered multiple automated jobs that fetched the same code. The earlier setup struggled with slow transfers, failed merges, and reliability that sometimes fell below 90%. OpenAI reports that median push latency fell by nearly 80%.
The system now delivers nearly a petabyte of Git data daily and has served more than 57 terabytes of compressed Git packs in one hour. Those figures are company-reported production numbers, not independent benchmarks.
Branch Storage and Download Delivery
Pando handles updates to each branch separately, so unrelated pushes can proceed in parallel. Incoming packfiles, which are Git’s compressed bundles of objects, go into blob storage, and branch versions and their dependencies are recorded in a separate metadata store. That design avoids forcing every branch update through a single repository-wide coordination point.
Updates stay ordered within each branch, but readers who fetch several branches can see versions from different moments. Pando therefore does not provide a single, point-in-time view of every branch across the repository. For downloads, it reuses stored packs instead of building a custom response for every client, and its Express delivery service uses regional caches.
OpenAI reports cache hit rates above 99%. It also says tests on standard cloud virtual machines saturated 10 Gbps links. Related client improvements are available through OpenAI’s public Git repository.
Security Testing Runs Alongside Deployment
OpenAI implemented Pando in Rust without legacy dependencies. Engineers built tests that compare its behavior with standard Git, along with repeatable simulations that check system rules and expose race conditions. Those checks targeted bugs in transfer negotiation and in the selection of packs needed to reconstruct repository history.
Agents also used Lean and TLA+, tools for expressing and checking system properties, to help find implementation bugs. Some continuous testing sessions lasted hundreds of hours. Separately, OpenAI says agents using Daybreak-capable models continuously red-team Git and Pando, and the report does not give a vulnerability count or detailed findings from that work.
The report connects this approach with Codex Security’s vulnerability validation workflow and OpenAI’s broader Defense Factory security model, both aimed at investigating flaws and preparing fixes. Pando’s rollout began as a proxy to the old backend so migration could happen gradually without downtime. The account shows how faster AI-assisted development can be paired with continuous testing while human teams still handled design and deployment, and it does not establish that agent-built infrastructure is free from vulnerabilities.
Licensed under CC BY-ND 4.0 International