Why Jujutsu May Be Better Than Git for AI Coding Agents

Summary

Yuka Ooka argues that Jujutsu can be a better version-control system for AI coding agents even though Git has stronger first-class agent support. In the author's workflow, an agent starts each task in a new Jujutsu change, while automatic snapshots record edits without requiring a final commit or a staging-index ceremony. Because the working copy is also the current point in history, people and agents generally track fewer states than they do in Git. The article says this reduces cognitive overhead and avoids interruptions about whether changes should be committed or stashed before history operations. Ooka also presents Jujutsu's recovery model as a major advantage when agents make mistakes. Every Jujutsu operation is recorded in an operation log, and `jj undo` and `jj redo` can reverse both an operation and a recovery attempt. The evolution log and stable change IDs can help locate earlier revisions and recover from problems such as failed conflict resolutions. By contrast, Git distributes state across the working tree, index, stash, commits, and reflog, while tools such as Claude Code's `/rewind` may restore files without restoring Git's state or changes made outside the agent. The article further argues that clean, semantically coherent history matters more when developers read less AI-generated code: it helps agents retrieve relevant context, supports review, and speeds incident investigation. Jujutsu's change and bookmark model also simplifies splitting work and building dependent branches for stacked pull requests, although the author notes that agents need configuration because Jujutsu lacks Git's default support. The piece is an author's argument, not a comparative benchmark, and concludes that tools designed to reduce human cognitive load can also make AI-controlled development safer and more stable.