There are some apps like this that leverage git-worktrees under the hood: https://conductor.build
I've tried lazygit as well to make it more convenient.
I still end up preferring having multiple clones of my repo when I need to ensure the agents don't accidentally overlap.
Curious what other people do.
Depth first search via recursive dispatch is my current go-to strategy. This requires a fully serialized chain of operations. Not even tool calls are allowed to run in parallel. One agent and one action at a time. That's it.
This does feel really slow at first, but in practice it seems to converge on high quality solutions more quickly than breadth first techniques which leverage parallelism and more tokens. The solution itself might not be the most ideal, but we find it very quickly and with a great deal of consistency. This means we can iterate more rapidly and with more certainty.
Just because we can consume $100 of tokens in 60 seconds doesn't mean we should try to. I think trying to go fast is what's burning a lot of people out on AI. There's a ton of value in here if we can slow down and be a little bit more deliberate about it.
First of all, the limiting factor is my attention. Even when I really tried to do this in parallel I could not meaningfully run more than three sessions.
I realized that some of the attention is devoted to thinking how multiple agents could step on each others toes as and end up creating merge conflicts when it came time to merge.
I have multiple projects and if I feel I have capacity to do parallel coding, I do it in unrelated code bases.
Sometimes I do super async work - as in one interaction per hour or so. I do it while walking, waiting in line, or between meetings. I found web based Codex to be decent for this. I guess it’s in parallel with my Claude Code sessions.