OpenAI rebuilt its Git backend (Pando) because of coding-agent traffic, says its CTO of Applications

By Nate (Nate's Substack) · Fri, 6 Nov 2026.

The email is the introduction (public section) of an article by Nate based on an interview with Vijaye Raji, OpenAI's CTO of Applications. Raji joined OpenAI through the acquisition of Statsig, the company he founded.

According to the text, OpenAI's coding agents increased the amount of work flowing through its development systems. Engineers can pursue more changes at once, and agents repeatedly fetch code, run tests and push their work. The infrastructure supporting those activities struggled to keep up.

OpenAI built Pando to solve the repository side of the problem. The company published its technical write-up on November 2. The article's outline mentions two design changes behind Pando and a drop of nearly 80% in median push latency, a figure Nate says OpenAI reports. The email does not specify what those two changes are.

The author's central thesis is that when one part of the work speeds up, it changes what the rest of the organization has to absorb: each individual engineer is more productive, but the team delivers much less of that gain. He proposes identifying the resulting delays, understanding their causes and changing the systems or decisions responsible. He adds that agents can also help with that work.

The full article's outline includes five sections:

1. Why OpenAI rebuilt its Git backend: what agent traffic did to OpenAI's monorepo, Pando's two design changes and the nearly 80% drop in median push latency.

2. Where acceleration stalls: the case of FurtherAI, where merged changes per engineer quintupled and the bottleneck shifted to testing, review and deciding what to build.

3. Review rules an agent can run on its own: how FurtherAI turned recurring review comments into 21 automated checks, whose failures agents fix before a human steps in.

4. Prototypes you're willing to throw away: Vijaye's argument for not getting attached to code and how to present a working demo to a team without shutting down the debate.

5. A seven-question constraint review: a table to fill in with recent work, pinpoint where work is waiting and run an experiment.

The full conversation with Vijaye accompanies the piece, and the complete content, along with Nate's Library MCP and his Slack community, is reserved for paid subscribers. The email ends there, so no further detail is available on Pando or the interview.