Git AI Everywhere

Last 4th of July, Sasha and I made the first commits to Git AI. At the time, we weren’t sure if AI was actually making us more productive — or just making us feel more productive.

So we built a Git extension to track AI-generated code through the full SDLC — from prompt to production. We started mining the data: which agents worked best, how to give them the right context, how engineers actually collaborate with them, and which tasks are still worth doing by hand.

The data was mixed for a while.

Then, in late November, something changed.

Teams using Git AI started seeing sharp increases in:

• local acceptance rates
• % of AI-generated code in each PR
• the durability of AI-generated code after it shipped

That was the Opus 4.5 moment. Right there in the data.

Now there’s a more general acceptance that Coding Agents have the potential to make engineering orgs more effective and productive, but between engineers, teams, and even repos at the same company results seriously vary. Every engineering team is trying to figure out how to unlock the benefits of AI without breaking their products and muddying their codebases.

Some teams call this initiative “adopting AI” and others “building their software factory”. I’ve never been one for naming trends in-progress, but one thing is clear: our profession is forever changed and every engineering organization is trying to figure out how to best incorporate Coding Agents into their craft.

There’s a lot of loud opinions about what works - we’re here to give them real data to figure out what actually does.

Maximum Compatibility

Teams want Git AI running everywhere — transparently, without changing workflows or slowing down Git.

Earlier this week, we shipped version 1.2. We got a whole lot closer.

Git AI no longer requires wrapping Git or setting up hooks.

You still get accurate, line-level attribution that survives real-world operations — cherry-picks, rebases, squashes, resets, stash/pops — without running anything in Git’s hot path.

Instead, Git AI runs asynchronously, reconciling checkpoints and commits in the background without blocking your Git commands.

Prior versions of Git AI worked by wrapping Git. We met teams who were comfortable rolling this out, but we never liked being in the business of wrapping Git because even a small hit to performance can be felt by developers. We did ship a beta version of Git AI that only used Git hooks, but in practice they performed significantly worse due to the overhead of Git’s hook execution layer. At enterprise scale, setting up hooks in every repo is a manual, operational burden. Bad idea.

It’s been a lot of work, but we’re pretty close to our North Star now. You can install Git AI on a machine and immediately start getting accurate attribution across every repo, without changing your workflow or hook configuration.

Run git ai upgrade or install Git AI fresh from GitHub to try it out today.

Agent Support

Each week we get PRs that add support for new Agents to Git AI. Currently, there are open PRs for Kimi, Firebender, Kilo, Pi and Roo. We’ve been a bit backed-up with the big 1.2 release, but we're getting back to testing and merging support for new agents. Keep them coming!

You’ll also notice the first integrations for using Git AI with Background Agents went live. Many of our customers have been investing in moving more work to the background so they can parallelize tasks, and run agents more securely. When developers are not monitoring what the agents are doing, observability becomes very important. We’ll continue investing here and working with the agent companies to expand support.

And last thing - because it’s really cool. Some teams, inspired by Ramp's great string of blog posts and Uber’s X posts, have been using Git AI to observe and iterate on their own background agents. We think this is going to become more and more common this year and look forward to helping platform teams get the data they need to build great Background Agents.

Thanks to all the contributors who are bringing Git AI everywhere, and to everyone else in the community who brings issues, product feedback and ideas. It’s been a lot of fun. We are thankful to be working on this problem with all of you and to be on the same journey – figuring out how to effectively bring AI into everything we do.