Quick answer
OpenAI outgrew Git because AI agents write and check out code far faster than tools built for offline humans. In Nate B Jones’s interview, Vijaye Raji (CTO of Applications at OpenAI, via the Statsig acquisition) explains that after February 2026, parallel Codex agents drove a surge in pull requests. Reviews, CI/CD and the monorepo itself became the bottleneck — so OpenAI built an internal repository system for always-on agents. For everyone else, the lesson is practical: if coding gets cheap, fix reviews, tests and release gates next.
Key takeaways
- February 2026 is when Raji says powerful models plus parallel Codex agents changed OpenAI’s shipping pace.
- Token usage at frontier labs grew about 17× in a year versus about 2× at average software shops, per the interview.
- More prototypes are thrown away because ideas become code quickly — and review, CI/CD and testing become the long pole.
- Git assumed offline work. Agents are online all the time, so checkout throughput and sparse checkouts matter as much as merge collisions.
- Founder litmus test: if a new model drop scares you, you may be wrapping the model; if it excites you, you are building on top of it.
- Your action: measure review and CI wait time, not only “lines of code generated”.
In this guide
- What broke when OpenAI’s agents got faster?
- Why was Git the wrong assumption set?
- Which bottlenecks show up after coding gets cheap?
- How did OpenAI solve the repository problem?
- What does the next billion ChatGPT users need?
- How are leaders using Codex to set direction?
- What is the founder litmus test for new models?
- How did the ads team reach a billion-dollar run rate?
- What should a small team copy from this?
- What does the interview not confirm?
- Watch: the source video
- Frequently asked questions
- The bottom line
- Related posts
Faster coding sounds like pure upside. Inside a large lab, it exposes every slow step after the first draft. Nate B Jones sat down with Vijaye Raji to unpack what broke, what OpenAI rebuilt, and what leaders should copy.
Credit for the interview goes to Nate B Jones (@NateBJones); the judgements below are ours. The full conversation sits on Nate’s Substack for paid subscribers.
What broke when OpenAI’s agents got faster?
Raji points to a combination of stronger models and agentic coding with multiple agents in parallel, including worktrees. Engineer PR volume jumped. Token usage is a blunt proxy, but he says frontier labs saw about a 17× rise in a year while typical shops saw about 2×.
Two effects follow: teams generate more prototypes (and throw more away), and the old conversation-heavy planning step stops being the long pole. Reviews, CI/CD, testing and instrumentation take its place.
🛠️ Practical guides
Ready to put this into practice? These step-by-step guides answer two common follow-up questions:
- How to Fix Code Reviews and Testing When AI Writes More Code — shorten review and CI queues when AI multiplies pull requests
- The Founder Litmus Test for New AI Models — check whether your product gets better when frontier models improve
Why was Git the wrong assumption set?
Git was designed for developers who check out code, work offline, then push. Agents do not work that way. They stay online, pull constantly, run tests and push back. Raji says OpenAI’s monorepo traffic rose by more than an order of magnitude versus what the old system was built for.
💡 Tip: You do not need to replace Git tomorrow. You do need to notice whether checkout time, lock contention or flaky CI is eating the gains from AI coding.
Which bottlenecks show up after coding gets cheap?
| Bottleneck | What shows up | First fix to try |
|---|---|---|
| Code review | Humans cannot read 5× more diffs with the same process | Auto-review for low-risk changes; smaller PRs; clearer ownership |
| CI/CD | Queues grow; flaky tests block agents and people | Split critical vs nice-to-have checks; cache aggressively |
| Checkout / monorepo | Agents hammer main; full clones are too heavy | Sparse checkouts, shared caches, read replicas |
| Production quality | More code reaches continuous deployment | Stronger gates on high-risk paths; better observability |
Raji notes that merge collisions on main are real, but checkout throughput was the bigger pain: agents need sparse checkouts and caches so they are not all hitting main for full trees.
How did OpenAI solve the repository problem?
OpenAI stood up an internal team with a mission, not a prescribed architecture: solve the repository problem for agentic load while keeping existing work running. The result, per Raji, is an internally built system that now supports the monorepo and the engineers on it.
The meta-lesson Nate draws: AI adoption creates second- and third-order tooling debt. Hand a small team the problem and the constraints, not a slide deck of “how”.
What does the next billion ChatGPT users need?
ChatGPT has about 1.2 billion users, Raji says — largely through a text box. The next billion are less likely to be comfortable typing English. After time in India, he argues voice is the modality that matches how many people already use WhatsApp: voice notes, not keyboards.
How are leaders using Codex to set direction?
Instead of only pitching ideas to product teams, Raji prototypes with Codex against the monorepo — for example, a voice-first ChatGPT shell that opens the mic immediately and replays what was said once the connection is ready. Prototypes do not ship as-is; they change the quality of the conversation with the team.
Feeling the review backlog already? Use our guide on fixing reviews and testing when AI writes more code before you buy another coding seat.
What is the founder litmus test for new models?
Raji’s litmus test for founders: models will keep getting better. If each new model drop makes you nervous, you are probably competing with the model. If each drop makes your product better for free, you are using models the right way. Our short guide on the founder litmus test turns that into a checklist.
How did the ads team reach a billion-dollar run rate?
OpenAI’s ads team, about nine to ten months old at the time of the talk, reached about a billion-dollar ARR run rate in roughly seven to eight months from zero code, Raji says — with a team on the order of 50 people. They leaned on LLM capabilities for rewriting, ranking and related pieces instead of training a custom model for every stage. As base models improve, the ads stack gets smarter without a full rewrite.
⚠️ Warning: That ads figure is an internal claim from the interview. It is not audited public financials. Use it as a story about leverage, not as a benchmark for your own revenue plan.
What should a small team copy from this?
- Measure the queue after codegen — review lag, CI minutes, flaky-test rate.
- Add auto-review only where risk is low, and keep humans on auth, payments and data deletion paths.
- Invest in sparse checkouts and caches if agents clone large repos all day.
- Prototype strategy with agents, then let the owning team productionise.
- Apply the litmus test before you build a thin wrapper around today’s model.
What does the interview not confirm?
🔎 Unconfirmed / rumour: Token-usage multiples, ChatGPT’s 1.2B users, ads ARR, team size and the exact design of OpenAI’s internal repository are as stated by Raji in the interview. We have not verified them, and internal tooling is not a product you can buy.
Watch: the source video
What this video covers: Nate B Jones interviews Vijaye Raji on how OpenAI’s coding agents overloaded Git-era repositories, what bottlenecks appeared next, how leaders prototype with Codex, the founder litmus test for model drops, and how the ads team used LLMs to move fast.
Video: Nate B Jones (@NateBJones), interviewing Vijaye Raji, CTO of Applications at OpenAI. This summary and the opinions in it are our own. Internal OpenAI figures (token usage, ads ARR, user counts) are as stated in the interview and have not been independently confirmed.
Frequently asked questions
Did OpenAI literally delete Git?
The interview says they built an internal repository system designed for always-on agents because Git’s offline assumptions did not fit. It does not mean Git is obsolete for every team.
What should a ten-person startup do first?
Do not rebuild Git. Fix review SLAs, CI time and test reliability so AI-written code does not pile up unreviewed.
What is the founder litmus test?
If a new frontier model makes your product better automatically, you are building on models. If it threatens to replace you, you may be wrapping them. See our litmus test guide.
Is voice really the path to the next billion users?
It is Raji’s view based on how people already communicate in markets where typing English is awkward. Validate with your own users.
Where is the full interview?
Nate points paid Substack subscribers to the full conversation and deep dive; links are in the YouTube description.
The bottom line
When agents make code cheap, the constraint moves downstream. OpenAI’s answer was a new repository stack and faster reviews; your answer might be simpler: shorter PRs, stronger CI and a product that gets better when models improve. Speed without gates is just a faster way to ship debt.
Put it into practice: tighten reviews and testing for AI-written code, then run the founder litmus test on your roadmap.
Related posts
- Is an AI Agent’s Cost Worth It?
- Agentic Control Plane: Choosing AI Agent Platforms
- AI Agent Project Readiness Brief
