On November 2, OpenAI engineers described Pando, their internal system for storing and serving Git repositories. The company says Pando now serves its main shared repository. It aims to let many people and agents push changes to different branches concurrently while keeping the usual Git commands.
There are two sides to this problem. Writes must accept concurrent changes, while reads must deliver large amounts of data during clones and updates. Faster writes alone do not make clones faster. Successful updates to one branch also do not eliminate conflicts between changes to the same branch.
OpenAI reports a nearly fivefold drop in median push latency in its own environment. This is a company result for a specific workload. Another team would also want to measure its slowest requests, failure behavior and initial clone time. Pando illustrates how agent workloads can change the requirements for familiar development tools.
*Cover: conceptual AI illustration.*
