Hacker News Reader

Comment by skrebbel | Hacker News Reader

Vaguely related question, for people who run multiple agents in parallel: how do you do that? do they all work on the same files at the same time? won't that create a terrible mess? or do you like check out the codebase 8 times into separate directories and do a complicated PR flow type thing?

View on HN

Some people use git worktrees so each agent is working in a separate directory.

Some people have multiple agents working in the same directory, and allow the agents to tell each other what they're doing and which files they're editing: https://github.com/Dicklesworthstone/mcp_agent_mail

you can use https://git-scm.com/docs/git-worktree to have many "views" of a single git repo.

For more complex environments, use multiple VMs (common to have large VMs running in AWS/GCP for development and only use your laptop or phone as a thin client)

I don't follow, so these people using 8 claude codes in parallel, they have 8 git worktrees? and 8 devservers listening on 8 ports?

or 8 VMs? won't syncing be super slow there? like i believe it, but it feels like a lot of setup and i don't understand why nobody seems to be talking about it.

> but it feels like a lot of setup

Ah, but you can have Claude Code set it up for you!

I am also curious about this, how people wrap their head around multiple changes in multiple features at the same time?

I personally, can handle only one feature at a time, prompt the AI, refine the changes, re-plan, review changes, rewrite parts when necessary, but my mind is still with one feature.

How do you folks manage multiple parallel feature development at the same time?

> how people wrap their head around multiple changes in multiple features at the same time?

Long before AI I routinely had 5 parallel development environments I would flip between. (I do distributed systems stuff, think HPC-adjacent clusters, when you're too big for Kubernetes and too small for supercomputing). One would be running long-running tests for my current changes, one being reset to a clean state, and 3 more on standby so I had 100% "duty cycle" (the reset took a while). You got used to multitasking.

Its different though, because your tasks were long-running and you are forced to have idle time between those runs.

In case of AI, it just delivers things in 2-3 minutes, this time is not enough (or worth) to switch the context (for me personally)

Git worktrees for parallel repo clones on different branches, a worktree setup script that runs when a worktree is created (npm install etc) and vibe-kanban [0] to visually create, monitor and approve parallel work

0 https://www.vibekanban.com/

Reply on news.ycombinator.com