OpenAI built Pando, a custom Git-compatible backend, because standard Git cannot handle the workload Codex generates. Codex runs thousands of parallel coding agents, each requiring its own isolated repository clone. A single monorepo cloned thousands of times simultaneously breaks conventional Git infrastructure at the object storage and network layers.
Pando solves this by virtualizing repository state rather than duplicating it. Agents get a logical clone that shares underlying object storage, eliminating redundant data transfer. The system maintains full Git protocol compatibility, meaning existing tooling works without modification. The architectural details of how Pando manages object deduplication, ref namespacing, and concurrent write isolation are exactly what makes this paper worth reading in full.
This matters beyond OpenAI. Any organization running large fleets of CI agents, autonomous coding systems, or parallel development workflows hits the same Git scaling ceiling. Pando is a blueprint for what Git infrastructure looks like when the client count shifts from hundreds to hundreds of thousands.
[READ ORIGINAL →]