OpenAI has rolled out “Pando,” a freshly built Git backend, over a 10-week engineering sprint. It replaces the company’s previous shared code repository infrastructure, doing so without any pauses in development workflows. Pando now underpins all development activity in OpenAI’s monorepo, handling the growing strain from intense AI powered engineering efforts. Production traffic began routing through the system in April 2026.
How Pando Redesigns Git Infrastructure
Pando retains the Git protocol and familiar workflows for developers, but reworks the storage and delivery structure behind the scenes. It separates out updates by branch, allowing parallel pushes and writes instead of funneling everything through a single coordination bottleneck. Incoming packfiles — Git’s compressed bundles — are stored in blob storage, while branch versions and inter-dependencies are tracked in a distinct metadata store. While consistency is maintained within each individual branch, read operations spanning branches may see data from slightly different points in time.
The system also reuses stored packfiles rather than generating custom responses for every download. An Express delivery layer adds regional caching, with hit rates reportedly exceeding 99%. OpenAI cites that, in practice, the system delivers nearly a petabyte of Git data daily, and has served over 57 terabytes of compressed packfiles in a single hour under load. Median latency for push operations dropped by roughly 80%. These are internal production figures, not independently benchmarked.
Continuous Security: Agents, Tests, and Red-Teaming
Pando was built in Rust and avoids legacy dependencies. OpenAI engineers created tests that compare its behavior against standard Git, ran long simulations to detect race conditions and validate system rules, and focused on edge cases like transfer negotiation or pack selection during history reconstruction. Formal methods tools — including Lean and TLA+ — were employed to express invariants and surface implementation bugs. Some of these verification sessions stretched over hundreds of hours.
On top of that, OpenAI uses AI agents, especially those tied to its Daybreak-capable models, to continuously red-team both Git and Pando. These agents aim to locate vulnerabilities even as workflows scale. This effort slots into OpenAI’s broader security frameworks, including its vulnerability validation workflow and the so-called Defense Factory model. While the technical report describes this relentless scrutiny, it does not quantify how many vulnerabilities were found or how severe any were.
Pando was introduced gradually: initial deployment acted as a proxy to the previous system, enabling a seamless migration with no downtime. Human oversight remained central; one engineer outlined the core architecture, teams of agents implemented most of it, and a small engineering group handled production rollout. Continuous AI-assisted development was coupled with round-the-clock security testing to ensure reliability under pressure.
What this means: Pando represents a major scaling milestone for OpenAI’s engineering operations. The firm dealt with version-control bottlenecks—slow merges, failed pushes, and reliability dipping below 90% under load—by reinventing infrastructure from the ground up. Now, with Git performance much faster and delivery systems more efficient, OpenAI is better placed to sustain rapid growth.
Analysis: By combining AI agents, formal verification tools, and production-grade engineering, OpenAI is embracing a model where infrastructure itself is developed partly through autonomous systems. Pando doesn’t simply solve immediate performance issues; it signals a shift toward AI-assisted core infrastructure development. Key risks remain: without transparency on vulnerabilities unearthed during red-teams and tests, it’s hard to fully assess security robustness. What to watch: whether this model becomes a standard across AI firms, how findings from continuous testing are shared (internally or publicly), and whether agent-built components introduce new classes of maintainability or compliance challenges.
