Vercel says autonomous agents now author 25-35% of merged PRs on its AI SDK

Vercel has built an autonomous agent pipeline, internally called ai-sdk-factory, that now authors more than a quarter of the pull requests merged each week into the AI SDK, one of the most widely used open-source AI projects.

What's new

In an August 12, 2026 engineering post, Vercel wrote: "The AI SDK is one of the most popular open-source AI projects in the world," and disclosed that "between 25-35% of the PRs we merge on a weekly basis are now authored by" ai-sdk-factory agents. The system is triggered directly from the project's issue tracker: "GitHub webhooks feed the issue queue, and as soon as they arrive workers automatically pull them and kick off agent runs in sandboxes." From there, issues move through a sequential pipeline of agents — classification, analysis, implementation, and review — before anything reaches a human.

Vercel was explicit that the system is not fully autonomous end to end: "The last line of defense is human review: nothing is merged without approval from a human on the AI SDK team," and "a human stays in control throughout the entire process, including merging every change." To help reviewers calibrate how carefully to check a given change, the agents produce structured risk assessments that accompany each PR, letting human reviewers spend more scrutiny on higher-risk changes and less on routine ones.

Context

The AI SDK is Vercel's open-source toolkit for building AI-powered applications, and it sits at the center of Vercel's broader AI Gateway and agent-tooling strategy — the same week, the company also published usage data showing its AI Gateway routes tens of trillions of tokens monthly and introduced "Agent Plugins," an open specification for packaging agent skills and MCP servers. Running a high-traffic, widely depended-upon open-source project's own maintenance work through an agent pipeline is a notable dogfooding move: Vercel is using autonomous coding agents on the exact project developers use to build their own AI agents.

Why it matters

A 25-35% autonomous-PR rate on a major open-source project, with mandatory human merge approval preserved as the safety backstop, is a concrete, verifiable data point in the broader debate over how much of real software maintenance work can be safely delegated to AI agents today. It's a more grounded signal than vendor benchmark claims because it's Vercel's own production maintenance workflow on a project other developers depend on and can inspect the commit history of. The risk-calibrated review structure — agents self-assessing risk to help humans focus review effort — is also a pattern likely to get copied elsewhere as more engineering organizations experiment with agent-authored code at scale.