Which version-control tool should you give your coding agent? We pit Git, Jujutsu and GitButler against each other in a fight to the death
Designing tools for agents is a bit different than designing them for people. As we've all been using agents more to write, commit and share our code, we at GitButler have been interested in making our CLI the best possible tool for agents to handle various version control tasks.
Recently our cofounder Kiril pitted Git, Jujutsu and GitButler against each other to try to accomplish a range of version control tasks at the hands of a few popular coding agents to see how GitButler compares.
TLDR: GitButler ran about 60% faster than git with roughly 80% fewer commands while Jujutsu ran slower than git.
In other words, if you're using Codex or Claude Code to selectively commit, split, squash, amend or otherwise do interesting version control things, GitButler does it faster and with fewer tokens than Git or Jujutsu.
You can check out our full report and the source code that we used to run it at vcbench.dev.
Here's every run, one dot each, across the five scenarios we tested:
agent runs:
Notice how the orange dots cluster to the left - that indicates faster runs, in other words, less time spent waiting for the agent to do your version control dirty work.
Methodology
What we wanted to do is see fundamentally how agents use these tools to do version control stuff, so no code is written during a run. They each start from the same state, get a plain text command about squashing or amending, and we see how long it takes and if it was right.
Two agents (Claude Code on claude-opus-4-8, Codex on gpt-5.5) times three toolsets times ten runs = 300 graded runs. A deterministic script grades only the final Git state - any command sequence that produces the right history passes.
Also, no cheating - raw git write commands are blocked in the Jujutsu and GitButler runs, so no quiet fallback to git rebase -i.
I know it's coming from us and we're one of the tools, but you really should try it out, it's often shockingly good at doing very complicated Git manipulation.
If you want to try it yourself, install the GitButler CLI then run but skill install to superpower your agent, then ask it to do something that you've never wanted to try because rebase -i has always intimidated you.
Scott Chacon is a co-founder of GitHub and GitButler, where he builds innovative tools for modern version control. He has authored Pro Git and spoken globally on Git and software collaboration.