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 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.
In its November 2 technical report, OpenAI said one engineer designed Pando’s core, teams of agents implemented it, and a small group of engineers deployed it. Production traffic began flowing through the system in April 2026. Pando now supports all development in OpenAI’s monorepo, where teams keep code together.
Codex increased the number of branches, pushes, merges, and downloads hitting OpenAI’s source control systems. One engineer could prepare several changes at once, with each push triggering multiple automated jobs that fetched the same code. The previous setup struggled, with slow transfers, failed merges, and reliability sometimes dropping below 90%.
Pando keeps the Git protocol and familiar workflows while changing how servers store and deliver data. OpenAI reports that median push latency fell by nearly 80%.
The system delivers nearly a petabyte of Git data daily and has served more than 57 terabytes of compressed Git packs in one hour. These are company-reported production figures, not independent benchmarks.
OpenAI Builds Git Infrastructure
Pando manages updates to each branch separately, letting unrelated pushes proceed in parallel. It stores incoming packfiles, Git’s compressed bundles of objects, in blob storage and records branch versions and their dependencies in a separate metadata store. This avoids making every branch update compete through one repository-wide coordination point.
The trade-off is important. Updates remain ordered within an individual branch, but readers fetching several branches may see versions from different moments. Pando therefore does not promise a single, point-in-time view of every branch across the repository.
For downloads, Pando reuses stored packs instead of building a custom response for every client. Its Express delivery service uses regional caches, with reported cache hit rates above 99%. OpenAI says tests on standard cloud virtual machines saturated 10 Gbps links. Related client improvements are available through OpenAI’s public Git repository.
Continuous Red-Team Testing and Security Checks
OpenAI implemented Pando in Rust without legacy dependencies. Engineers built tests that compare its behavior with standard Git, alongside repeatable simulations that check system rules and expose race conditions. These checks targeted bugs in transfer negotiation and 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. The report does not provide a vulnerability count or detailed findings from that work.
This approach connects with Codex Security’s vulnerability validation workflow and OpenAI’s broader Defense Factory security model, both focused on investigating flaws and preparing fixes.
Pando’s rollout began as a proxy to the old backend, allowing gradual migration without downtime. For security teams, the account illustrates how faster AI-assisted development can be paired with continuous testing, while retaining human design and deployment roles. It does not establish that agent-built infrastructure is free from vulnerabilities.
Give your SOC the alert context to stops incidents quickly. Integrate Threat Intel from 16K SOCs.
