OpenAI Introduces Pando to Scale Monorepo Git Backend for Agentic Workloads

OpenAI engineers Aaron Friel, Taylor Blau, and Ted Nyman announced Pando, a new Git repository backend created to handle high concurrent load on the company's monorepo driven by Codex and automated software agents. Over the prior year, parallel workflows generated by AI coding tools sharply increased the frequency of branches, pushes, merges, CI jobs, and fetches, occasionally reducing repository reliability below 90%. To resolve bottlenecks caused by centralized coordination without breaking existing developer workflows or client tools, OpenAI retained standard Git client and protocol compatibility while rethinking server constraints. Pando decouples branch updates by abandoning global point-in-time repository consistency in favor of causal consistency per reference. Pushes to unrelated branches proceed independently as immutable objects in blob storage and metadata updates, allowing conflict handling only when writers push to the exact same reference simultaneously. For reads, Pando reuses pre-generated packfiles instead of assembling custom packs for individual client states, trading minor redundant downloads for high server throughput. According to OpenAI, Pando is deployed across its monorepo, reducing median push latency by nearly 5x and delivering nearly a petabyte of Git data daily.

Key takeaways

• OpenAI developed Pando, a new backend architecture powering its internal monorepo to handle concurrent load driven by Codex and agentic development.
• Pando preserves Git protocol and client compatibility while moving from repository-wide consistency to causal consistency per branch reference.
• Packfiles are reused across read requests rather than generated uniquely per client, minimizing server CPU load and maximizing fetch throughput.
• OpenAI reports Pando reduced median push latency by nearly 5x in production while delivering nearly a petabyte of Git data per day.