Ask HN: How did you set up a multi-agent orchestration for personal use? I am about to start a bigger project with several subprojects and want to be a human in the loop of otherwise fairly automated loop, which should ideally use different agents for different roles (coding, design, testing, supervision, etc). I know this approach is becoming increasingly common and I wonder if there's already some tooling that facilitates that? Or do you simply use some specific model as a coordinator? |
OpenCode subagents
In my experience, I've gone back to basics and want to make them really good before going back up to multi-agent. This is why I'm using opencode now, one top-level agent, no message passing. Scion can then run these OC agent trees and they can send messages to each other. If you don't have the basics in a good place, they will go off the rails and you will burn more tokens.
There are some opencode plugins that take things further. Keeps things simple and understandable first, then move up the stack. (imo)
Yes: the subagent couldn't talk to the parent anymore after a first return to it, but that didn't matter. That I could beget new useful interact able agents was amazing.
I'm mad and bitter angry and won't ever stop being mad until opencode or I fix it so that opencode really has multiple agents that can work with each other, and that I can work with.
I started doing some work on message passing mailbox systems, alike many of the other great https://github.com/joelhooks/swarm-tools or https://github.com/Dicklesworthstone/mcp_agent_mail_rust efforts , but super trailed off.
I have high interest in scion too! I love the robust platform underneath. Yes please. I've spent a lot of time digging into the opencode data model underneath, & it's a moving inplementation defined system. Making the pieces of the system concrete like Scion has feels like a superpower.
Scion OC support is not good on main, I have a (outdated) fork which vastly improves it. There are a number of OC plugins that try to add higher level orchestration as well as independent wrappers like https://kubeopencode.github.io/kubeopencode/docs/features/
It's debatable if you need multiple harness support like Scion or something more focussed like KubeOpenCode
Checked out swarm tools for a few seconds until the main pages of the docs 404'd, not a good sign when the first three menu items do that...
One cool thing about Scion, you can attach directly to the agents in the terminal or the web ui, even seeing real-time typing appear across them if you have both open (thanks tmux!). Scion also handles all the messiness of setting up a container with auth, git worktree, and tmux. I tried to get a simple version going but it became a rabbit hole of complexity and edge cases. You can run it in "workstation" mode which is all local.
One thing I keep getting stuck on with these systems though: they all seem to assume an agent lives in one worktree of one git repo. Maybe that is fine in a monorepo world. Outside of that, the repo boundary is often just not the task boundary. Some context lives next door, or two repos away, and the sandbox somehow has to know what to bring in.
I'm wondering if there is a tool in the agent orchestration space that prepares multi-repository worktrees for a subagent out of the box?
There's some share .scion space that allegedly your orchestrator / team is I think supposed to be able to share with other tasks, I think... Not sure. No idea how it works. But that's at least some cross worker scratch space... Again allegedly.
You might be able to ask an agent, use a skill, or have it done on demand.
I think there are hooks coming that we can tie into. Open an issue on GH, ptone is good about replying, just not timely all the time.