OpenAI Unveils Pando, a Git Backend Built for AI Coding Agents
OpenAI has introduced Pando, a Git-compatible backend that handles about a petabyte of data a day and cuts push latency almost fivefold under heavy AI-driven commit loads.
SUMO· the number puzzle
Dialo· learn languages by speaking
ediviewer· read EDI files like documents
giftdetect· the right gift in minutesOpenAI has unveiled Pando, its own backend for the Git version control system, designed for a world in which AI assistants produce most of the commits. The company says Pando stays fully compatible with standard Git tools, processes roughly a petabyte of Git data per day and reduces push latency by almost five times.
Why the old servers buckled
The trigger was Codex and similar coding tools. According to OpenAI, quarterly commits to its monorepo will grow by more than 2,400% by the end of the year. Traditional Git servers were built for human engineers making dozens or hundreds of changes a day; AI assistants push that to thousands of commits and merges per hour. Reliability of the earlier infrastructure fell below 90%, with slow clones and downloads, merge errors and limits on repository size.
OpenAI chose to keep the Git ecosystem rather than replace it, concluding that the bottleneck was the architecture of conventional servers, not the client or protocol.
How Pando works
Building Pando took 10 weeks. One engineer designed the architecture, and groups of Codex agents wrote the implementation in Rust. The system went into production in April.
Pando drops two classic requirements: a single consistent state across all branches, and a custom packfile built for every client request. Each branch is written independently, so a push never waits on unrelated branches. Incoming pack files are validated and passed straight to blob storage without being unpacked, while metadata records the new branch tip, file location and dependencies. For new branches, only the trunk is advertised as a base, which keeps ref lists small. The trade-off is that repeated pushes to the same branch resend whatever is not yet in trunk, which OpenAI considers acceptable because the median pack file in its monorepo is under 20 KB.
For fetches, Pando hands clients links to stored packs via the packfile-uri mechanism. A caching service called Express serves them from regional caches with a hit rate above 99%, according to the company, and clones reach the 10 Gbit/s network limit on standard virtual machines. Background jobs merge small packs into larger ones and nothing is deleted, which lowers CPU and disk I/O load.
Availability
OpenAI also fixed several issues in the Git client itself, and some changes have been accepted upstream, including parallel indexing and an explicit haves feature. The Pando backend is not open source. Announced on November 2, 2026, it is in limited preview for selected enterprise partners. Wider commercial access is planned for the first quarter of 2027, priced by repository size and operation count, with monitoring for AI-generated code share, merge conflict rates and repository growth.
Apps worth a look
SUMO: the number puzzleTap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.
Dialo: learn languages by speakingTalk to an AI conversation partner; every sentence comes back corrected and explained in your language. 9 languages, A1 to C2.
ediviewer: read EDI files like documentsOpen EDIFACT and X12 messages, see every segment explained, convert to JSON/XML, export PDF and Excel. Runs locally, no upload.
giftdetect: the right gift in minutesPick who it is for, the occasion and their personality. The search engine reads product descriptions from local shops and ranks what fits. Free, no sign-up.
More from this day
See all →
GreenShift accelerator graduates 12 Egyptian green scaleups
Journify raises $12M Series A at four times its valuation
Cogniguard's Vguard Stimulates the Vagus Nerve During Sleep for Alzheimer's
