OpenAI has introduced Pando, an innovative backend for Git, developed in a mere 10 weeks utilizing Codex AI agents. This new system replaces the existing infrastructure supporting OpenAI’s primary code repository, achieving this transition without disrupting ongoing engineering operations.
Continuous Security and Performance Enhancements
In their technical report dated November 2, OpenAI detailed the creation of Pando, spearheaded by a single engineer, with additional support from AI agents and a small engineering team for deployment. Production traffic started using Pando in April 2026, enhancing OpenAI’s monorepo, a centralized repository for code management.
The introduction of Codex has significantly increased operational efficiency, managing more branches, pushes, and merges than before. Previous systems faced issues with slow transfers and unreliable merges, sometimes dropping below a 90% success rate.
Redefining Data Management and Delivery
While retaining the Git protocol and traditional workflows, Pando revolutionizes data storage and delivery. OpenAI reports a remarkable 80% reduction in median push latency, showing substantial performance improvements.
Pando processes approximately a petabyte of Git data daily and has handled over 57 terabytes of compressed Git packs within an hour, according to OpenAI’s internal data. This data management shift allows for independent updates per branch, reducing conflicts and improving efficiency.
Advanced Security Measures and Testing Protocols
OpenAI developed Pando using Rust, eliminating legacy dependencies, and implemented rigorous testing protocols. These tests, including simulations and comparisons with standard Git behavior, focus on maintaining system integrity and identifying potential bugs.
AI agents utilize Lean and TLA+ to express and verify system properties, uncovering bugs that traditional methods might miss. Continuous testing sometimes extends over hundreds of hours, with AI models persistently red-teaming both Git and Pando to identify and mitigate vulnerabilities.
The gradual rollout of Pando as a proxy for the old system facilitated a seamless transition without service interruptions. This implementation highlights how AI-driven development can align with continuous testing, maintaining the balance between innovative automation and essential human oversight.
The development of Pando underscores OpenAI’s commitment to advancing AI-assisted infrastructure while ensuring robust security protocols are in place. This effort represents a significant step forward in leveraging AI for software development efficiency and reliability.
