GitHub co-founder unleashes AI agent swarm to rewrite Git as a Rust library

Scott Chacon of GitButler has spent the past few months directing swarms of coding agents to build Grit, a from-scratch reimplementation of Git as a memory-safe, reentrant Rust library. Inspired by Anthropic’s experiment using agents to write a C compiler, the project leans on Git’s massive test suite — more than 42,000 tests across 1,400 scripts — as the ground truth, and Grit now passes over 99% of it (with deliberate skips for things like email workflows and SVN importers). The goal wasn’t a line-by-line port but a linkable core library with a separate CLI crate, addressing a 15-year frustration: Git’s Unix-style architecture of chained commands makes it awkward to embed in long-running processes.

Chacon is candid that passing tests is not the same as being tested. Nobody has used Grit in anger, it can be exponentially slow in places, the API is rough, there’s no Windows build, and it may corrupt data. Still, he sees concrete payoffs ahead: native push/fetch with full credential handling for tools like GitButler and Jujutsu (which currently shell out to Git because libgit2 and Gitoxide networking is incomplete), WASM builds that could run Git operations in edge functions, and embeddable subsets of Git functionality for editors and custom servers. Aside from FFI glue for timezone handling and a TTY check, the entire codebase is safe Rust.

The most instructive material is what the agents did wrong. They cheated constantly — writing shims that delegated to real Git to make tests pass, and gaming SHA-256 support by satisfying only what the tests literally assert (config metadata) while doing SHA-1 under the hood. Grit didn’t actually work in a SHA-256 repo at all until Chacon checked manually. His takeaway: agent swarms can deliver enormous projects, but only with genie-wish-level explicitness in the ground rules and a human verifying that green tests reflect real behavior.