The new system has reduced code-push latency nearly fivefold, revealing how autonomous coding is changing the infrastructure required to manage large software projects.
OpenAI has developed a new Git-compatible source-control backend called Pando after rapid growth in AI-assisted software development overwhelmed parts of its existing infrastructure.
The company disclosed the system in a November 2 engineering report authored by Aaron Friel, Taylor Blau and Ted Nyman, members of OpenAI’s technical staff.
According to OpenAI, Pando now supports development across its main code repository, reducing median push latency by nearly five times while serving approximately one petabyte of Git data daily.
The announcement highlights an emerging challenge for software engineering: as AI coding agents generate and modify more code, infrastructure designed around traditional human development workflows may struggle to handle the increased activity.
AI Coding Creates New Infrastructure Bottlenecks
OpenAI said Codex has enabled engineers to perform substantially more development work in parallel.
That expansion produces more branches, commits, merges and automated testing operations, increasing pressure on source-control infrastructure.
According to The New Stack’s technical analysis, OpenAI’s quarterly code commits could increase by more than 2,400% by the end of 2026, compared with the first quarter of 2024.
That figure is a company projection, not a completed annual result.
OpenAI reported that its previous infrastructure experienced slow fetches, delayed pushes and failed merges, with reliability occasionally falling below 90%.
The company responded by redesigning how Git data is stored and delivered rather than abandoning the Git ecosystem.
Pando Changes How Code Is Stored and Shared
Pando retains compatibility with standard Git clients, allowing developers and AI agents to continue using familiar commands.
Its architecture changes how repository updates are coordinated and how compressed code packages, known as packfiles, are distributed.
Rather than requiring every server to maintain an identical view of all branches simultaneously, Pando uses causal consistency to manage updates more independently.
According to OpenAI’s engineering explanation, the system also reuses stored packfiles across client requests instead of repeatedly generating separate copies.
These decisions reduce unnecessary coordination and data processing, although they introduce trade-offs in how different branches are observed.
OpenAI built the system in Rust over approximately 10 weeks, with engineers designing the architecture and Codex agents contributing implementation and testing.
Pando began handling production traffic in April 2026.
Anthropic Is Pursuing a Different Architecture
OpenAI is not alone in reconsidering source control.
In an October engineering report, Anthropic described building Scion, an internal version-control system designed around AI agents that remain connected to centralized infrastructure.
Anthropic reported that its average daily merged changes increased from approximately 2,500 in May to more than 30,000 by September 2026.
Unlike OpenAI, which preserved Git compatibility, Anthropic migrated its main repository away from Git.
The contrasting approaches illustrate different infrastructure strategies: adapting an established ecosystem or replacing it with technology designed for extremely high-volume automated development.
What This Means for Software Developers
For engineering teams, the announcements suggest that AI-assisted coding may eventually shift infrastructure bottlenecks from code generation toward version control, testing and integration.
The official Git protocol documentation describes the existing mechanisms for exchanging compressed repository data between clients and servers. Scaling those operations becomes increasingly important as automated systems generate more concurrent requests.
However, OpenAI and Anthropic represent unusually large development environments. Their experiences do not establish that ordinary engineering teams need to replace existing source-control systems.
For founders and infrastructure providers, the bigger opportunity is identifying which parts of the software development process become expensive as AI-generated changes multiply.
The next competitive challenge in developer tools may not be generating code faster, but managing, validating and integrating that code reliably at scale.
