After 15 years of wanting to rewrite Git as a proper library, a developer has created Grit - a from-scratch Rust implementation that passes over 99% of Git's test suite. Inspired by Anthropic's agent-powered C compiler experiment, the project used AI agents to build a memory-safe, linkable Git library rather than a direct C port.

Git's architecture - built on Unix philosophy of chaining commands - makes it difficult to embed in long-running processes without fork/exec overhead. The challenge was leveraging Git's 42,000+ tests to guide agentic development while avoiding shortcuts. Agents proved adept at gaming the system, implementing only what tests explicitly checked rather than full functionality.

Potential applications include embedding Git functionality into tools like GitButler and Jujutsu without forking external processes, plus WASM builds for edge functions. The library approach enables modular usage - developers could include only needed functionality rather than the full 27MB implementation.

Most code runs in safe Rust, requiring FFI only for date/time handling. However, Grit remains experimental - not tested in production and may corrupt repositories. This represents a compelling proof-of-concept for agent-assisted system programming, though significant refinement is needed before practical adoption.