AI code agents flood open source with 10x more pull requests no one can review

Code agents are working. Jensen Huang says we’ve gone from 30 million coders to a billion overnight. Anyone with an agent can spot an open issue on GitHub, generate a fix, and submit a PR. It feels like contributing. But for projects like Hugging Face’s transformers library—over a billion downloads, thousands of projects depending on it—this is less a gift and more a siege.

PR volume has jumped tenfold. The number of maintainers hasn’t. These aren’t people who can rubber-stamp code. Transformers is built as a human-to-human communication channel through code that reads top to bottom, favoring flat hierarchies so practitioners can follow the logic without untangling abstractions. Agents don’t know this. They don’t know that a design decision exists because no one wrote it down. So they submit “improvements” that follow generic best practices, break implicit contracts, introduce subtle bugs, and refactor things that weren’t broken. They’re also sycophantic—they’ll diligently implement any suggestion, including ideas a seasoned maintainer would have killed with a three-word comment.

This is the new open-source bottleneck. It’s not about accepting or rejecting AI code. It’s about the human effort required just to read, understand, and evaluate what an agent produced in seconds. Every PR still needs a maintainer to check design direction, spot side effects, and write feedback. That math doesn’t work when submissions multiply but humans don’t. The same pressure is hitting Apple’s App Store reviewers, who are drowning in a flood of agent-built apps.

The team behind this work saw an opportunity in the relationship between transformers and MLX, Apple’s machine learning framework. Most mlx-lm models are ported from transformers implementations, which have become the de facto source of truth because of their emphasis on readability. That’s a clean, scoped problem for an agent: don’t invent from scratch, use the transformers code as a reference. They built a Skill that, given a prompt like “convert the olmo_hybrid architecture to MLX,” sets up an environment, downloads models from the Hub, reads the transformers code, writes an MLX implementation, runs tests, and iterates until results match. It won’t declare success until it’s actually correct.

The twist is how they’re positioning it. The Skill isn’t just for contributors—it’s an aide for reviewers too. It generates artifacts that increase signal: generation examples, numerical comparisons, and a separate non-agentic test harness for reproducibility. The goal is PRs that could have come from a careful human, with enough supporting evidence that a reviewer can evaluate them without exhaustively re-deriving everything. Their north star is simple: when a model lands in transformers, it should be available on MLX shortly after. Not through automation that overwhelms maintainers, but through tooling that respects them.