The evolution of AI-assisted programming

AI-enhanced programming has gone through several evolutionary stages in a very short time span:

1. First there were copilots that would answer questions about your code in a sidebar
2. The next big wave was command line tools like Claude Code, which allowed working on entire codebases using familiar command line tools

The next generation goes further: multiple agents working on the same codebase simultaneously, each isolated in its own git worktree. This has been a revolutionary change for many developers.

Multi-agent worktrees

Multi-agent tools with worktrees make it possible to quickly test different approaches to the same problem, or to work on several features in parallel -- while being able to turn back the clock at any point.

Once a thread of work is finished, the developer can review changes through diffs at any time. Code changes can even be reviewed by different AIs looking for bugs or other problems. When the change is ready, the AI generates a pull request from the discussion history -- with a detailed overview of what was done, what trade-offs were weighed, and what was decided.

That change can then be merged into the main branch -- or simply discarded. This way of programming is the future, hands down.