What happened
OpenAI's post "Scaling Git at OpenAI" describes Pando, a new backend for its monorepo. Every figure below is OpenAI's own, read through the outlets that reported it, because OpenAI's page would not load for our tools. Nobody independent has benchmarked any of this.
The trigger was load. Reliability fell below 90% "at times" across slow fetches, slow pushes, and failed merges, from around the end of 2025 (vendor, via Qodo). Quarterly commits are on course to grow over 2,400% by the end of 2026 against Q1 2024 (vendor, via The New Stack). Gigazine reports more than 14x by Q3 2026 (vendor, via Gigazine).
One engineer designed the core, Codex agents implemented it in Rust, and it reached production in April after 10 weeks (vendor, via The New Stack). Git clients and the protocol did not change.
| Measure | Reported result | Label |
|---|---|---|
| Median push latency | Nearly 5x lower, about 80% | vendor |
| Git data served | Nearly a petabyte a day | vendor |
| Peak delivery | Over 57 TB in one hour | vendor |
| Regional cache hit rate | Above 99% | vendor |
| 1,000 clients each pushing a commit | 7.71 s, against 56.52 s and 96.26 s for two other designs | vendor, Gigazine only |
What it gave up
Pando stops ordering branch updates across the whole repository. Pushes to different branches no longer wait on each other. The cost: a client reading several branches can see them at different moments, and OpenAI argues users only care about order within one branch (vendor, via The New Stack).
That is the part to read twice. If your release tooling reads several refs and assumes one point in time, this design breaks an assumption you never wrote down (Learnetto makes the same point, independent). Only trunk is advertised as a base too, so stacked branches resend everything outside trunk (vendor, via Qodo).
Two labs, two answers
OpenAI kept Git and replaced the server. Anthropic replaced Git itself for its monorepo with Scion, moving commits and queries server-side, and does not plan to release it (independent, Arize; The New Stack).
Neither is for sale. Both answer the same pressure from agent traffic, at volumes few enterprises have. Anthropic's Scion figures of over 30,000 merges a day (independent, Arize) describe a frontier lab, not your Salesforce DX repository.
Find your row
| What you see | Likely meaning | Action |
|---|---|---|
| Agent PRs up, review queue long, Git fine | The bottleneck is people | Fix review first, not Git |
| Slow clones and CI fetches | Object transfer | Try partial clone, shallow fetch, or sparse checkout |
| Same pack rebuilt for many CI jobs | Repeat server work | Cache immutable objects; ask your host about packfile-uri |
| Failed pushes or merges under concurrent agents | Write contention | Isolate agent branches, escalate to your host with data |
| All of it, on a monorepo you host yourself | Real backend limit | Scope a proper architecture review |
Five steps
Prerequisites: read access to your Git host's metrics and CI logs.
- Step 1. Separate the numbers. Track branch creation, pushes, fetches, merge failures, and CI fan-out apart from repository size. Owner: developer platform lead.
- Step 2. Tag agent traffic. Give agents their own identity so you can see their share. If you cannot, you are guessing. Owner: DevOps lead.
- Step 3. Set a reliability floor. Pick the merge and push success rate you will not go below, and alert on it. OpenAI's warning sign was under 90%, and that is far too late for you. Owner: engineering director.
- Step 4. Audit multi-ref readers. List every release and compliance tool that reads more than one branch. You need this before any host or vendor offers you relaxed consistency. Owner: release manager.
- Step 5. Ask your host. What does it support on packfile-uri and partial clone, and what is its concurrent push limit? Get the answer in writing. Owner: procurement with the platform lead.
Serving Git is your host's job. Knowing what your agents are doing to it is yours.
Trade-offs
- Tune what you have. Cheapest and reversible, and it covers most teams. It stops helping when writes, not transfers, are the constraint.
- Build your own backend. OpenAI's 10 weeks sits on Codex at OpenAI's disposal, its own metadata store, its Express cache, and its own Git build. Do not read 10 weeks as a plan.
- Wait for vendors. GitHub, Cursor, and Entire are all working here (independent, The New Stack). Waiting is reasonable while Step 1 is your only evidence. It is weak if you are already failing merges.
Missing
- OpenAI's post itself. It would not load for our tools, so we relied on The New Stack, Qodo, Gigazine, Cybersecurity News, and Learnetto. Check any figure against the original before you quote it.
- Any independent benchmark of Pando, and any cost comparison with a Git host.
- A release. Pando is not open source, and OpenAI says only that it is exploring ways to share the technology (via The New Stack). No migration path exists.
- Security results. OpenAI reports agents red-teaming the system (via Cybersecurity News), with no vulnerability count. Agent-built does not mean verified.
- A test by Prestanda. We have not run Pando, which nobody outside OpenAI can do, and we have no client measurement of agent Git load.
Frequently asked questions
Can we buy or download Pando?
No. OpenAI has not open-sourced it and publishes no API. Some of its Git client fixes are upstream, and others sit in OpenAI's public Git repository, as reported by The New Stack.
Do we need to replace Git for coding agents?
Not on this evidence. OpenAI kept the Git client and protocol and replaced the server. Measure pushes, fetches, merge failures, and CI fan-out first, because your constraint may be code review and not Git.
What is the risk in Pando's design?
It orders updates per branch, so a client reading several branches can see them at different moments. Any release or compliance tool that assumes one point in time across branches needs checking before you adopt a similar model.
Living page. Last updated 5 Nov 2026. We update this note when OpenAI releases any part of Pando or an independent benchmark appears.