Operations | Monitoring | ITSM | DevOps | Cloud

GitBench is our attempt to see how well LLMs handle git-specific tasks. We put together a suite of 17 benchmark categories with a total of 204 specific test cases. We ran these benchmarks against 43 (and counting) different models with all of their supported effort/reasoning levels.

This minor update has a one big new feature.

Since I started using Claude Code more frequently, I kept noticing a “worktree” checkbox popping up whenever I started a session in a Git repository. I had no idea what it meant, so I did what any curious developer would do and started digging. What I found was a Git feature I somehow never came across before: git worktrees.

From plan to main. Software is no longer just a tool. It is the infrastructure of modern life. Software keeps airplanes in the sky and power flowing into our homes. It helps doctors save lives, scientists discover cures, farmers feed cities, and astronauts navigate space. It powers economies, protects supply chains, and connects billions of people across the world. Every major system humanity depends on now depends on software. Which means developers are no longer just building applications.

You’re no longer writing code. You’re managing a pipeline of agents writing it for you. If you’ve been running two, three, or four AI coding agents in parallel, you already know the problem. The agents are fast. The orchestration is chaos. You’re bouncing between terminal windows, manually rebasing branches, cleaning up messy commits, and trying to remember which agent is touching which repo.