Vercel says its AI software factory now authors up to 35% of merged PRs

Vercel says the agent system it built to handle work on the AI SDK has moved beyond experiments and into measurable repo output. In an August 12 post on X, the company said its software factory now authors up to 35% of merged pull requests, closed 70% of issues in July, and helped bring open bugs down by 25%.

The official source largely backs that up. In Vercel's blog post, Building a software factory for AI SDK, the company says the system autonomously processes issues and pull requests while humans stay in control of every merge. Vercel describes the last four weeks as a period where the factory authored 25 to 40 percent of merged PRs, which lines up with the 35 percent figure highlighted on X.

Vercel's write-up also gives more detail on how the system works. Instead of using a single general-purpose agent, the company says it split the workflow into specialized steps for classification, analysis, implementation, and review. The post also points to the surrounding stack, including Vercel Sandbox for agent workspaces, Neon Postgres for factory data, and public GitHub pull request history tied to the ai-sdk-factory bot.

That combination of concrete numbers and a public workflow is a big reason the story is getting attention on X. Plenty of companies talk about AI coding agents in broad terms, but far fewer publish repo-level metrics that suggest those agents are already affecting merged output and issue velocity in a live open source project.

For developers and product teams, the interesting part is not just that Vercel built an internal agent pipeline. It is that the company is framing software delivery as a managed production system where the job becomes improving the factory itself: tightening prompts, adding evals, reducing failure cases, and deciding which steps still need a human in the loop. If that model holds up, teams may start measuring agent contribution the same way they measure CI reliability or deployment throughput.

There are still important unknowns. Vercel has not shown yet whether the same approach will generalize outside the AI SDK, how much operator time it takes to maintain the workflow, or how the system performs when issue mix and repo complexity change. The published numbers are strong, but they come from one codebase and one company that is unusually deep in developer tooling.

Even with those caveats, this is one of the clearer signals so far that coding-agent workflows are shifting from demo material to operational metrics. Vercel is not just saying agents can write code. It is saying they are already changing the shape of work on one of its flagship open source projects.