OpenAI scales Git for the AI Agent Era with the new Pando backend

The automation of code writing by artificial intelligence has pushed the limits of traditional software engineering infrastructure. On November 2, 2026, OpenAI officially announced Pando, a new version control backend built specifically to handle the massive workload generated by its engineers and the autonomous agents based on the Codex model. Developed in a period of just 10 weeks by technical team members Aaron Friel, Taylor Blau, and Ted Nyman, the system maintains full compatibility with the standard Git protocol and client, but radically alters how data is processed and distributed on the company's servers.

Pando focuses on solving two fundamental problems of Git scalability: horizontally scalable writing, optimizing simultaneous pushes to multiple branches without concurrent locks, and maximizing read throughput, saturating network bandwidth through intensive use of packfile caches. The transition to the new infrastructure occurred without the need for work interruptions or changes in development team workflows.

Background and context: OpenAI's monorepo crisis

Traditional software development using Git was conceived in the 2000s by Linus Torvalds, initially designed for a community exchanging operating system kernel patches via email. Before 2025, the workflow within OpenAI operated within the normal limits expected for large-scale repositories.

However, the introduction of the Codex model and the proliferation of AI agents changed the equation. A single engineer went from having a few tasks to multiple changes ready for continuous integration (CI) simultaneously, with each push triggering dozens of parallel jobs that needed to fetch the same commit. Automation multiplied data traffic in the repository.

The chronology of OpenAI's infrastructure collapse and reconstruction occurred as follows:

• Pre-2025: The development workflow followed industry standards, supported by conventional Git backends.
• Late 2025: The previous system collapsed due to severe bottlenecks. Critical slowness occurred in fetches and pushes, as well as systemic failures in merge processes. System reliability plummeted, falling below 90% at peak times.
• Q1 2024 to Q4 2026: A statistical explosion in quarterly commit volume was recorded, with a cumulative jump of 2,442%.
• Q3 2026: Faced with the technical limit imposed by the growth of autonomous agents, OpenAI's technical team developed Pando in a 10-week window, replacing the underlying infrastructure.
• November 2, 2026: OpenAI makes Pando's architecture and implementation data public.

Technical architecture and concrete data

Traditional Git synchronizes repositories by managing "objects" (blobs, trees, commits, and tags) and "references" (refs, such as branches and tags). When multiple users (or thousands of AI agents) attempt to update refs simultaneously, standard systems that impose a single order of recording through methodologies such as Three-phase commit (3PC) or Write-ahead logs (WAL) in blob storage create bottlenecks. One push ends up waiting for another, even on unrelated branches.

Pando bypassed this architectural limitation by coordinating updates for each branch completely independently. Updates to unrelated branches proceed in parallel. For reads, instead of generating a compressed file (packfile) individually for each client requesting data, Pando reuses packfiles stored in a distributed cache (such as regional Express Caches), serving the same raw bytes to multiple clients with near-zero computational cost for the server.

The results reported by the Pando implementation are absolute in terms of data volume and performance:

• Daily volume: The system delivers nearly 1 petabyte of Git data per day.
• Traffic peak: In a period of high demand, the system served more than 57 terabytes of data in just one hour. More than 99% of these bytes were delivered directly from the cache.
• Reliability: The success rate in merges and system availability rose from the concerning <90% in 2025 to 99.99% currently.
• Latency reduction: Median push latency was reduced drastically. However, there is a slight technical divergence in the published metrics: while the official OpenAI source reports a cut of "almost 5x," external analyses report the reduction as "almost 80%."

It is vital to note that independent infrastructure researchers (as reported by Kingy.ai) highlight that all capacity and latency numbers are "first-hand" metrics. To date, there has been no independent reproduction or external stress testing to audit the numbers released by OpenAI.

Impact on the sector, users, and the Latin American market

The re-engineering of Git by OpenAI reverberates across multiple layers of the technology ecosystem, creating precedents for companies that integrate Artificial Intelligence at the foundation of their corporate engineering processes.

For OpenAI Engineers, the impact is direct and quantifiable: the end of merge locks and the acceleration of the CI/CD feedback cycle allow human work to keep pace with the production of code generated by machines, without interruptions from service outages.

For AI Agents (Codex), the change ensures the necessary infrastructure for continuous operation. Without Pando, agents were competing for network resources and database locks, resulting in timeouts that nullified the productivity gains of artificial intelligence. The new backend enables the execution of "swarms" of simultaneous autonomous agents testing code hypotheses in isolated sandboxes.

Regarding the Open Source Community, the impact will be progressive. The development of this technology exposes the historical "technical debt" of Git in the face of non-human workloads. As authors Friel, Blau, and Nyman made explicit, OpenAI plans to extract the performance improvements created in the internal client and send them (upstream) to the official Git project tree.

For the Latin American software market, the adoption of Pando serves as an early warning. Startups and unicorns in Brazil, Mexico, and Colombia that are implementing so-called AI coding factories (AI-guided software factories) will inevitably encounter the same architectural bottlenecks as they scale automated code submissions. Understanding how OpenAI fragmented ref coordination into independent updates and adopted aggressive packfile caching will guide the DevOps infrastructure of these companies over the next three years.

What is still unknown

Despite the wealth of operational details in the material released by OpenAI, substantial gaps persist in the widespread adoption of the technology:

• Licensing and Open Source: OpenAI's briefing and the statement from its engineers focus on sending specific patches back to the Git project (upstream). However, it is unclear whether the entirety of the Pando backend will be made available under an Open Source license or whether it will remain strictly a proprietary competitive advantage of the company's internal infrastructure.
• Exhaustive Architecture: The exact internal structure of blob storage, the format of hierarchical covers, and the management of push metadata in O(1) complexity were described in broad mathematical terms, but lack the exhaustive detail necessary for autonomous replication by third-party systems architects.
• Third-Party Testing: Empirical verification of Pando's performance under adverse workloads that are not specifically those of the OpenAI monorepo has not yet been documented.

What to expect next

The practical timeline for version control technologies derived from the Pando case points to the following concrete actions in the coming quarters of 2027:

• Upstream Patch Integration: The Linux community and the maintainers of the Git project should receive specific pull requests originating from OpenAI's internal fork (ttaylorr/git), with an immediate focus on packfile negotiation efficiency in the transfer protocol and metadata read optimization in O(log Δ) complexity.
• Evolution of DevOps Tooling: Companies like GitHub (Microsoft) and GitLab will inevitably respond to the documentation released by OpenAI by adapting concepts from Pando (such as Express Cache) into their offerings aimed at heavy corporate instances, preparing the commercial ground for the native integration of AI agents into B2B client workflows.

Frequently Asked Questions

Why did OpenAI's version control fail in late 2025? The old system was designed for human operations. With the intense introduction of the Codex model and multiple autonomous agents developing code in parallel, traffic jumped 2,442%, creating bottlenecks in fetches and failures in merges, dropping system reliability to less than 90%.

What does Pando do differently from traditional Git? While traditional Git coordinates all repositories by imposing a strict linear order or requiring approval from multiple nodes (as in Three-phase commit), Pando allows writing to multiple branches simultaneously without one interrupting the other and reuses compressed files (packfiles) on cache servers, freeing up computational power.

Do engineers need to relearn how to use Pando system commands? No. Pando operates exclusively on the backend (on the servers). It was built to maintain 100% compatibility with the standard Git protocol and client, so developers and agents continue using the same commands and conventional tools of their routines without noticing the structural change.