When a company's coding agents start opening branches and running jobs in parallel at scale, the quiet part of the toolchain that no one thinks about, source control, becomes the thing that breaks. OpenAI says that is exactly what happened to its own Git service, and its response was to build a new backend from scratch. The useful signal here is not the tool itself but the load pattern, which any team scaling up agents will meet in their own continuous integration and version control.
The pressure came from Codex, which let OpenAI's engineers do far more work at once: more branches, more pushes, more fetches from development environments, research jobs and CI. Starting late last year the systems behind its monorepo struggled, with reliability at times dropping below 90%, meaning slow or failed pushes and merges. Rather than abandon Git, OpenAI kept the familiar workflow and rebuilt the backend underneath it, naming the result Pando. The design goals were to scale writes across many branches at once and to serve the same history to many readers from cache.
The reported results are substantial. Pando cut the latency of a median push by nearly 80%, and now serves close to a petabyte of Git data a day, with more than 57 terabytes in a single hour and over 99% of those bytes coming from local disk cache. OpenAI built it in Rust with no legacy dependencies, and says it went from kickoff to production in about 10 weeks, with Codex itself and a mix of models doing much of the work. It began serving production traffic in April.
The relevance to the rest of us is the pattern, not the product. Pando is not open source, so most teams cannot adopt it. Only the client-side improvements OpenAI contributed back to Git are public, through its own fork, which means this reads more as a warning than a tool: the systems around Git were shaped by two decades of human workflows, and swarms of agents push them in ways those tools were not sized for.
The honest caveats are the usual ones for a vendor engineering post. The reliability and latency figures are OpenAI's own and have not been independently verified, and a backend tuned for one very large monorepo may not translate to smaller or differently shaped repositories. Still, as more teams point agents at their codebases, the lesson that parallel machine work strains CI and source control in new ways is worth taking seriously before it shows up as failed merges of your own. We review OpenAI Codex, the agent that created this load in the first place.