129,822 Commits in 3 Years

Matthew Diakonov··2 min read

129,822 Commits in 3 Years

Three years ago, the repo averaged 4 commits per day. Last month, it averaged 356. That is an 89x increase. The code did not get 89 times more complex. The tooling changed.

The Evolution

Phase one was Codex. Basic code completion. You wrote the structure, the model filled in boilerplate. Commits increased because you spent less time on repetitive code. Maybe a 3x improvement.

Phase two was GPT-4. Longer context, better reasoning. The model could understand entire functions and generate working implementations from descriptions. Commits doubled again because the unit of work shifted from lines to functions.

Phase three was Claude with tool use. The model could read files, run code, check errors, and iterate. A single prompt could produce a working feature across multiple files. Commits went exponential because the model was doing the edit-compile-test cycle autonomously.

Phase four is Opus with desktop integration. The model operates the entire development environment. It writes code, runs tests, checks the browser, fixes issues, and commits - all in a loop. The human role shifted from writing code to reviewing and directing.

What 356 Commits Per Day Means

Most of these commits are small. A function change here, a test fix there, a config update. The granularity increased because the cost of committing decreased. When a model is making changes, committing after each successful change is free.

The interesting metric is not commits per day but working features per week. That went from roughly two to roughly twelve. A 6x improvement in actual output, hidden inside an 89x increase in commit volume.

The Quality Question

More commits does not mean better code. It means more iterations. The code is better because each change is tested before the next one starts. The model catches its own errors in the next cycle. Rapid iteration with verification produces cleaner code than careful single-pass implementation.

Fazm is an open source macOS AI agent. Open source on GitHub.

More on This Topic

Related Posts