OpenAI has deployed the Pando Git backend to manage a steep rise in automated code submissions across its primary software repository, cutting median push latency by nearly 80 per cent. The custom system serves close to one petabyte of Git data every day across the organisation.
Key points
- OpenAI built Pando in Rust over 10 weeks to cope with a 2,442 per cent increase in quarterly commits driven by Codex agents.
- The architecture isolates branch updates rather than coordinating across the whole repository, boosting merge reliability to 99.99 per cent.
- A regional caching layer called Express delivers cache hit rates above 99 per cent and allows fetches to saturate 10 Gbps network connections.
- OpenAI is contributing several client-side transfer patches back to the core open-source Git project, though Pando itself remains proprietary.
Handling commit surges from autonomous agents
The rapid integration of autonomous coding systems across OpenAI created unexpected bottlenecks for traditional engineering tooling. As internal teams paired human staff with autonomous Codex agents, development velocity accelerated sharply. By the fourth quarter of 2026, quarterly commits inside the company’s central monorepo had grown by 2,442 per cent compared with the first quarter of 2024.
Conventional Git infrastructure struggled to accommodate this workload. Repository reliability slipped below 90 per cent as automated test suites and agent jobs repeatedly launched concurrent downloads, merge attempts, and branch updates. Slow network transfers and failed operations threatened to stall day-to-day engineering output.
The issue is spreading across the software sector as generative tools produce higher volumes of code. As more companies consult guides like the AI agents roadmap to deploy automated coding systems, repository traffic continues to surge. Other platforms are addressing similar pressures, with GitHub improving its infrastructure to handle billions of pull requests every month. While some competitors have opted to build entirely separate version control systems, OpenAI chose to preserve compatibility with standard Git commands.
How Pando alters repository coordination
Writing in the project announcement, OpenAI engineers Aaron Friel, Taylor Blau, and Ted Nyman asked if the organisation had “reached a limit of Git itself, or of the way we served our monorepo?” The team determined that the core protocol remained sound, but standard server architectures created unnecessary delays.
Traditional Git servers require extensive coordination when clients push changes. In standard set-ups, the server maintains a strict, consistent sequence across every branch at once. If hundreds of automated agents push code to unrelated branches at the same moment, their updates queue behind one another while the server locks the repository to calculate global state.
Pando drops this requirement by adopting causal consistency. Instead of locking the whole repository, Pando isolates updates on a per-branch basis. Moving a branch pointer to a newer commit requires coordination only for that individual branch. Pushes to independent feature branches can proceed in parallel without waiting for concurrent writes elsewhere.
This design introduces a specific technical trade-off. A client inspecting multiple branches at the exact same instant might view them at slightly different points in time. OpenAI determined that this behaviour causes no friction in practice, as automated agents and human developers almost exclusively track changes on one branch at a time. The change restored operational stability, lifting merge reliability from below 90 per cent to 99.99 per cent.
Packfile streaming and regional caching
Pando also alters the way Git servers store and deliver repository objects. When a developer pushes code, a standard Git server uncompresses the objects and unpacks them into a local file system. When another client fetches that branch, the server recalculates missing objects and compresses a fresh packfile on the fly. Generating these tailored bundles consumes heavy processor time and storage bandwidth.
Before an upload, a regular server also advertises every branch it holds so the client can omit objects the server already stores. With millions of branches, that reference list alone can exceed 100 megabytes. Pando avoids this overhead by advertising only the primary trunk branch as a base, keeping the reference list tiny. While repeat pushes to the same branch must resend any intermediate history not yet merged into trunk, OpenAI noted that the median pack in its monorepo remains under 20 kilobytes.
Instead of unpacking incoming commits, Pando streams new packfiles directly into cloud object storage. It logs the location of the packfile, its dependencies, and the target branch commit in an independent metadata store. Because each pack builds directly on the advertised trunk, the stored files remain reusable by any client that already holds trunk history.
To distribute these bundles quickly, OpenAI built a delivery service called Express. Express stores packfiles in regional caches near developer clusters and test runners, achieving a cache hit rate above 99 per cent. In benchmarks on standard cloud virtual machines, client downloads saturated 10 Gbps network connections. During peak traffic periods, Pando has processed more than 57 terabytes of Git packfiles within a single hour.
Standard Git servers regularly run garbage collection and repacking cycles to clean up loose objects, consuming heavy compute and disk resources. Pando never discards packs. Instead, background worker services continuously merge smaller packages into larger groups, operating much like a log-structured merge database. This compaction allows the server to locate required packfiles in logarithmic time while retaining every compaction layer.
Open source contributions and industry rivals
OpenAI developed Pando in Rust, completing the initial system in 10 weeks. A single systems engineer designed the architecture, while teams of Codex agents handled implementation and built testing tools to identify concurrency bugs. The software entered production in April 2026. To keep developer routines unchanged, engineers continue to run standard Git commands, supported by an internal Git build distributed to staff machines and packaged inside the ChatGPT desktop application.
While Pando itself remains closed source, OpenAI is contributing several transport fixes back to the public Git project. Pando relies heavily on the packfile-uri protocol feature, which a Google engineer first proposed in 2019 to allow clients to download packs over separate HTTP connections.
During internal testing, OpenAI engineers resolved issues with concurrent downloads, improved cleanup routines after dropped connections, and added resumption support for interrupted transfers. OpenAI also introduced detailed tracing mechanisms to distinguish time spent generating packages from time spent transferring network bytes. Other patches hosted in the company’s public repository include parallel pack downloading and an explicit haves feature to streamline branch negotiation.
The engineering challenge has produced diverging strategies across the technology industry. Startups such as Entire, founded by former GitHub chief executive Thomas Dohmke, and Cursor with its Origin project, have also built on top of Git. Competitors like Anthropic have chosen to abandon Git compatibility altogether, building a proprietary client and architecture called Scion from scratch. Like Anthropic, OpenAI has kept its backend private, though it stated that it is exploring ways to share the underlying design with external developers.
