Daily · Article 04 · 5 min read
- OpenAI
OpenAI explains Pando: where code from parallel AI work goes
Technical announcement dated November 2 · original read November 4 JST
Summary
In a November 2 technical article, OpenAI introduced its internal code-delivery system Pando. Its parallel transfers, human and AI roles, and nearly 80% reduction in push latency need distinct interpretations.
When AI produces several changes at once, the place that collects and tests them gets busier too. In a technical article dated November 2, OpenAI introduced Pando, its internal system for exchanging code history. This is infrastructure for sending and receiving work, rather than a model that writes code faster.
Key facts
AI developer OpenAI says Pando now supports development in its monorepo and remains compatible with Git tools and protocols. Git records versions of files and their history; a monorepo manages code for multiple development efforts together. This explanation reads the November 2-dated original on November 4 JST. OpenAI engineering announcement
The company reports reducing median push latency by nearly 80% in production. That is its own result, not a measurement of every development step or outside users' work improving equally. It says it is exploring ways to share Pando's technology with users.
1. Storing files and running them
Copies named “final,” “final revised” and “really final” make it difficult to know which version was tested. Git version control records a file set's state and the history before it. A commit records one point; a branch is a name pointing to a line of work. Separate branches let different changes proceed before deciding what to combine. Git's version-control explanation · How branches work
This is a fictional example of two changes handled separately. Sending code to a repository does not itself release a service.
| Work | What the repository retains | Decision that follows |
|---|---|---|
| AI changes a search screen | New commit on a search branch | Test search and review the change |
| Person changes payment wording | New commit on a wording branch | Review language and user flow |
| Combine the results | Commit containing the combined code | Test the combination and decide on release |
- Work
AI changes a search screen
- What the repository retains
- New commit on a search branch
- Decision that follows
- Test search and review the change
- Work
Person changes payment wording
- What the repository retains
- New commit on a wording branch
- Decision that follows
- Review language and user flow
- Work
Combine the results
- What the repository retains
- Commit containing the combined code
- Decision that follows
- Test the combination and decide on release
The cat asks Q&A
The researcherIf both jobs edit the same file, does the repository choose the correct answer?
Pocket NotebookCorrectness depends on the work itself. Git supplies history and conflict information, but the combined result still needs testing. Saving separate branches quickly and deciding whether changes are correct are separate checks.
2. Bottlenecks when sending and receiving
A push sends your change history to a shared repository. Fetch and clone bring needed history into another working environment. Continuous integration, or CI, also needs the code before automated tests can start. More simultaneous AI changes multiply both incoming submissions and jobs downloading the same revision.
OpenAI describes handling updates to unrelated branches independently and reusing compressed packs for multiple readers. A pack is a file containing history data for transfer. Caches retain the same pack on nearby servers, reducing the work of packaging the same material repeatedly.
Reads of different branches cannot be assumed to show one exact moment across the whole repository. Simultaneous changes to the same branch still need coordination. The design delivers required history while reducing waits between unrelated work.
%% The changed code-delivery path · simplified from OpenAI's Pando description
flowchart TD
A["Developer and AI jobs<br/>Send changes on separate branches"] --> B["Pando<br/>Independent updates across branches"]
B --> C["Stored packs and caches<br/>Reuse for multiple readers"]
C --> D["Other environments and tests<br/>Receive needed code history"]The cat asks Q&A
- The researcher
Is a branch name enough to record when I want to revisit a test result?
Pocket NotebookRecord the tested commit identifier too. Later changes can move the branch name to another point. Keeping the tested revision makes it easier to compare the result with the same code later.
3. What people designed and agents implemented
OpenAI says the project went from kickoff to production in 10 weeks. One engineer designed the core, teams of agents implemented it, and a handful of engineers deployed it. Reducing that account to “AI built the system alone” removes design and operating decisions.
The company describes tests against standard Git behavior and repeatable simulations that expose errors caused by operations occurring in difficult orders. This is its account of development; this publication has not installed or tested Pando.
| Role | Work described in the announcement | Result to distinguish |
|---|---|---|
| Engineers | Core design and production deployment | Decide the structure and where to apply it |
| Agent teams | Implementation and participation in validation | Produce code and test material for review |
| Validation system | Git-conformance and reproducible simulation tests | Check existing behavior and concurrent-work errors |
- Role
Engineers
- Work described in the announcement
- Core design and production deployment
- Result to distinguish
- Decide the structure and where to apply it
- Role
Agent teams
- Work described in the announcement
- Implementation and participation in validation
- Result to distinguish
- Produce code and test material for review
- Role
Validation system
- Work described in the announcement
- Git-conformance and reproducible simulation tests
- Result to distinguish
- Check existing behavior and concurrent-work errors
The cat asks Q&A
- The researcher
If AI also writes tests, who decides whether to trust the result?
Pocket NotebookThe operating team must define required behavior and failure conditions, then review the evidence. Records using different inputs and operation orders, or comparing standard Git, help that judgment. What was compared matters more than a bare “tests passed” statement.
4. The time behind “nearly 80%”
The reported metric is median push latency. A median is the middle value in an ordered set of measurements, not the longest wait experienced by any request.
A development team increasing AI work can first separate its waiting stages. Writing code, transfer queues, test queues and review all take different time. Faster repository transfers can leave testing or review as the next bottleneck.
The public OpenAI Git repository contains improvements to Git itself. It should not be read as the complete source of the internal Pando service or a hosting product anyone can use today. The earlier HackerTube explainer also separates human preparation, AI work and evidence of completion in a real web project.
This normalization explains the nearly 80% reduction on one common basis; it is not a measurement in seconds.
| Same push-latency basis | Relative time | How to read it |
|---|---|---|
| Before | 100 | Chosen comparison baseline |
| After | About 20 | 100 × (1 − 0.80), about one fifth |
| Work excluded | Not calculated by this table | Writing, testing, review and service deployment |
- Same push-latency basis
Before
- Relative time
- 100
- How to read it
- Chosen comparison baseline
- Same push-latency basis
After
- Relative time
- About 20
- How to read it
- 100 × (1 − 0.80), about one fifth
- Same push-latency basis
Work excluded
- Relative time
- Not calculated by this table
- How to read it
- Writing, testing, review and service deployment
The cat asks Q&A
- The researcher
Should a small team switch to Pando straight away?
Pocket NotebookFirst find whether the wait is in code transfers, tests or review. These results come from OpenAI's large repository, and the article does not announce a product with established outside-access terms. Connect your own bottleneck to tools you can actually use first.
To close
What comes after faster code writing
More AI jobs make the channels that store and distribute their results busier. Pando's case shows the operating work of designing code delivery, testing and review as one flow alongside the model's capabilities.
Sources
- OpenAI engineering announcementopenai.com
- Git's version-control explanationgit-scm.com
- How branches workgit-scm.com
- OpenAI Git repositorygithub.com