OpenAI built Pando infrastructure in 10 weeks to handle a 2,442% code commit surge

Summary

OpenAI developed Pando, a custom Git infrastructure, in 10 weeks with the help of Codex to handle a 2,442% quarterly increase in commits triggered by AI agents. This reduced push latency by approximately 80% and improved merge reliability to 99.99%.

Why it matters

This is a rare concrete example of AI agents generating so much development traffic that it overwhelmed the existing infrastructure, necessitating a bespoke solution. It also highlights a self-reinforcing cycle: the agents helped build the tooling that now serves them, compressing the design-production loop into 10 weeks. For every organization scaling agentic coding, the architectural choices (branch-based causal consistency, packfile reuse, regional caching) offer directly transferable lessons.

Highlights

• A 2,442% increase in commits, driven largely by Codex agents rather than just human engineers (2024 Q1 → 2026 Q4)
• Causal consistency design: Branches coordinate independently instead of sharing a global write-ahead log, enabling concurrent updates without a single bottleneck
• Developed in Rust over 10 weeks by a team of architects and agents; over 99% of 57 TB/hour of data is served from local cache

Critical take

The 2,442% headline appears inflated by the low base effect from 2024 Q1. The article relies almost entirely on a tweet from an OpenAI developer, reading more like a press release than independent reporting. The causal consistency trade-off (cross-branch reads may not be up-to-date) is dismissed in a single sentence and has not been tested in real-world scenarios. Consequently, the 99.99% reliability claim remains unverified by third parties.