OpenAI Pando shows the hidden infrastructure cost of coding agents
OpenAI rebuilt its Git backend for Codex-scale concurrency. Most teams should copy the measurements and controls, not the architecture.
The infrastructure signal
On November 2, OpenAI disclosed Pando, the Git-compatible backend now serving its monorepo. It says Codex-driven parallel work multiplied branches, pushes, merges and CI fetches until the previous system suffered slow operations, failed merges and reliability below 90% at times.
What Pando changes
Pando preserves the developer workflow while allowing unrelated branches to update independently and stored packfiles to be reused by many readers. OpenAI reports nearly fivefold lower median push latency and almost a petabyte of Git data served each day.
Our operating decision
Do not copy this architecture on reputation alone. Keep Git, measure branch and CI queueing, merge failures, fetch fan-out and permission risk, then isolate agent work and add evaluation and approval gates before increasing concurrency.
Our view is that Pando is not a cue for every engineering team to rebuild Git. It is evidence that once coding agents work in parallel, the bottleneck moves from model intelligence to the machinery around them: branch isolation, CI fan-out, repository reads, merge coordination, permissions and audit. OpenAI had an exceptional internal load and responded with a Git-compatible backend that cut its reported median push latency nearly fivefold, but Pando remains internal. We would keep the familiar Git interface, instrument queue time and failure rates across fetch, test and merge stages, cap agent concurrency by service ownership, and require traceable approvals for high-risk changes. Only when those measurements show source control itself is the constraint would we consider a backend migration. For most teams, copying Pando now would add a large platform burden while leaving the more common problems—conflicting changes, weak evaluation and unsafe permissions—untouched.
Turn insight into a working system.
We build agentic engineering workflows around your existing Git and CI, with scoped permissions, audit trails, evaluations and shadow runs before parallel agents touch production code. If your team is scaling coding agents and merge or CI pressure is becoming operational risk, it is worth talking to us.