OpenAI launches Pando to handle massive Git traffic from AI agents

The scale of AI-driven code

OpenAI recently introduced Pando, a specialized Git backend designed to power its internal monorepo. The company needed a new solution because the sheer volume of work performed by Codex agents began to overwhelm standard Git infrastructure. By late 2025, OpenAI reported that pushes and fetches were slowing down, merges were failing, and reliability occasionally dropped below 90%.

The growth in activity has been staggering. OpenAI's commit charts show that commits landing on their master branch increased by 1,378% in Q3 2026 compared to Q1 2024. The company estimates this growth will continue to accelerate, reaching an increase of over 2,442% by the end of 2026.

To solve these bottlenecks, the engineering team concluded that the issue was not with the Git protocol itself, but with how the server handled the data. Pando was built to manage these thousands of concurrent pushers without any degradation in performance. As marktechpost.com reported, the new system now delivers nearly a petabyte of Git data every single day.

Breaking Git's core assumptions

Pando achieves its speed by discarding two traditional requirements of most Git servers. The first is repository-wide consistency, where servers usually order every update across the entire repository. This often forces unrelated pushes to wait for one another. Pando replaces this with causal consistency, where each branch is serialized independently.

Under this new model, clients only face conflicts if they attempt to push to the exact same reference at the same time. The trade-off is that a user viewing multiple branches might see them at slightly different points in time. This change allows for massive parallelism, which is essential for handling the high volume of parallel work generated by AI agents.

The second change involves how the server handles packfiles, which are the files used to transfer data. Standard servers build a fresh pack for every fetch, but Pando stores every pushed pack as an immutable blob. It then uses a scalable metadata store to track versions, allowing for much faster reads and writes during the push process.

Optimizing for fast delivery

Managing a monorepo with millions of branches creates a unique challenge for

Efficient hardware and caching

Pando is designed to run on commodity hardware, making it highly efficient for large-scale operations. In standard cloud environments, the system is capable of saturating 10 Gbps links during clone and fetch operations. This ensures that high-speed data transfer remains consistent even as the volume of data increases.

To maintain these speeds, OpenAI uses a delivery service called Express to cache packs regionally. By utilizing consistent hashing, this service achieves hit rates above 99%. This architecture allows the system to serve data quickly to developers regardless of their geographic location.

Automated testing and partnerships

The development of Pando involved sophisticated automated testing to ensure reliability. Engineers utilized agents running GPT-6 Astra to write Lean and TLA+ specs, which are formal methods used to find bugs in complex systems. This rigorous approach helped the team reach production in a relatively short timeframe while maintaining high performance.

While the backend remains internal, OpenAI is exploring ways to share some of the underlying technology. Readers interested in the broader AI landscape can follow the project's updates on Twitter or join the community on the ML SubReddit. For those looking to promote a GitHub Repo, Hugging Face Page, or a product release, partnership opportunities are also available.

Advanced engineering and testing

To manage the massive volume of data, Pando uses background workers that group individual packs into geometrically larger covers. This process is similar to LSM-tree compaction, which helps the system organize data efficiently. Because these covers act as a skip list, the system can perform pack selection in logarithmic time, ensuring that the monorepo remains navigable even as it grows.

OpenAI also conducted a 1,000-writer simulation to test Pando against other designs like 3-phase-commit and write-ahead-logs. To ensure the system worked correctly, they wrote a Git conformance suite specifically for agents and added deterministic simulation testing. These rigorous methods allowed the team to verify the backend's reliability before it went live.

Deployment and community

OpenAI ships its own custom Git build to developer machines and the ChatGPT desktop app to support the new infrastructure. While the Pando backend remains internal, the company is exploring ways to share the technology in the future. They have also contributed several fixes for client handling of interrupted downloads and packfile-uri tracing to the upstream Git project.

For those interested in staying updated on these developments, you can follow the project on Twitter or join the community on the ML SubReddit. You can also subscribe to the newsletter for daily updates. If you are looking to partner for a GitHub Repo, Hugging Face Page, or a product release, the team is available for connections.