# The end of git? OpenAI CTO on the future of coding.

## Executive summary

OpenAI's CTO, Vijay Raji, discusses how the rapid deployment of AI-generated code, driven by agents, is fundamentally outgrowing legacy version control systems like Git. The increased velocity of code generation has shifted the primary bottlenecks from writing code to downstream processes, specifically code reviews, CI/CD, and the underlying code repository infrastructure. OpenAI has had to architect a new system designed for continuous, high-throughput, online agentic development.

## Key takeaways

- AI-driven code generation is outpacing existing systems.: The number of PRs put out by engineers has seen a massive uptick due to powerful models and agents running in parallel with work trees, leading to code generation speeds that exceed the capacity of traditional infrastructure.
- Downstream processes are the new bottlenecks.: As code becomes 'cheap' and prototypes are generated quickly, the critical bottlenecks shift to how teams perform code reviews, CI/CD, testing, and instrumentation on the increased volume of code.
- Git was designed for offline use cases, not continuous agents.: Because Git was built around an offline development story (checkout locally, then push), it cannot handle the constant, high-throughput, online nature of agentic development. OpenAI built a new system to address this.
- Scaling requires solving sparse checkout and main branch reconciliation.: The sheer volume of code going in and out of the mono-repo has exceeded the current system's design capacity. Solving the 'sparse checkout problem' and managing continuous deployments to `main` are critical scaling challenges.

## Technical details

- Code Repository Scaling: The current mono-repo structure at OpenAI is experiencing an order of magnitude increase in code flow, requiring a new repository system built with fundamentally different assumptions than Git's offline model. This new system is designed for the 'agentic world' with higher throughput.
- CI/CD Bottlenecks: The increased rate of code generation necessitates internal fixes for CI/CD infrastructure, including implementing systems for auto-reviews to keep pace with AI-assisted commits.
- Agentic Development Assumptions: The shift requires moving away from assumptions like 'main must remain main' and addressing technical issues such as the 'sparse checkout problem' and ensuring robust caching to prevent traffic issues hitting the main branch.
- LLM-Powered Efficiency: The Ads team achieved a $1B ARR run rate in under a year by leveraging LLM capabilities (e.g., ads indexer, query rewriter, ranker) in a way that was far more efficient than building a dedicated ML model for each component.

## Practical implications

- Build engineers must anticipate that AI agents will accelerate development to a point where traditional CI/CD and version control systems become the primary scaling bottlenecks.
- Architecting code repositories for continuous, high-throughput, online agentic workflows requires abandoning assumptions rooted in offline development models (like Git's original design).
- Focus on solving 'second and third order effects' (e.g., review process, repository throughput) rather than just the initial coding problem.
- The ability to rapidly prototype and validate ideas using LLMs (e.g., voice modality) is becoming a core capability for product development, democratizing the ability to build.

## Topics

AI Agents, Version Control Systems, CI/CD, Software Architecture, LLMs, Mono-repos, Substack Deep Dive

Source: https://www.youtube.com/watch?v=S6SEgMooZpE
