Cursor Engineer Merged 2,500 PRs in a Month: Where Did Human Review Go?

At a glance

Cursor's Lauren Tan said the team shipped 2,500 pull requests to production in a month, with no human review. That does not mean review disappeared. The work that review used to do has been split into several stages and moved elsewhere. The public pstack repository shows how this flow works.

Key points

1. 1

   Before opening a PR, an agent runs the app itself and leaves evidence that it works.

2. 2

   For every PR, a verification agent that did not write the code decides whether it passes.

3. 3

   When the same problem comes up again, it goes into code or automated checks instead of relying on people's memory.

4. 4

   Unpublished figures such as the number of PRs or incidents cannot be confirmed from the article.

Why it matters

As AI agents write a lot of code, the old way of having people read every line has limits. Where and how verification happens has become important.