LLM Codegen go Brrr – Parallelization with Git Worktrees and Tmux

7 points by skeptrune

alexkorban

What I don’t get is how this actually saves time. I’d have to review and test N versions of the code, and then likely re-prompt and review and test N versions again etc.? A toggle is perhaps a best case scenario - you just look at it and it’s obvious whether it’s working or not - but most code isn’t like that.

This setup is kind of predicated on total commitment to vibecoding, but in fact it’s really unclear that the current quality of agent output is good enough to save any time (even when babysitting just one agent), and so maybe it’s more productive to just write the code manually, once, and be done with it?

trickster

Also this might work for the first round, but the moment the solutions start to diverge then what? Are we supposed to abandon agents in a session?
- cdxker
  
  Exactly, the initial idea for now is to just insta abandon all the previous agents once you get one working the way you want.
skeptrune

I effectively paid $0.40 for someone else to make the wireframe for me. Definitely a time savings.

kaiwenwang_dot_me

True. Suspect this is the future. See these divisions so far:

Inline code edits
Whole file edits
Planning across multiple files and multiple changes
Pair programming/manually approving changes
Branching AI agent worktrees and selecting the best
AI performs pull request edits from issue tracker or PRD

alper

Git worktrees are pretty slow. Maybe just do jj new main instead?

satvikberi

How would this work in parallel? Wouldn’t each instance of claude code be operating on the same change?
danthegoodman1

Comment removed by author