I get that there's only so many good names, but Murmur is the name of the Mumble server implementation [https://www.mumble.info/], which is also communication. That makes this name collision worse than others.
Mind sharing your solution a bit more - how do they trigger actions from each other for instance (or is that prompted by you asking them to read from each other)? Asking because OP is interesting but seems a bit overambitious for my own usecases (two pairs of claude + codex agents such that each pair should talk between each other but not across pairs; all running on the same machine).
“You’re in a tmux pane, and the adjacent pane is another agent building a PR for me. Watch it and make sure it doesn’t get stuck before it opens the PR.”
It might make a tmux helper script, a tmux skill might help, but coming in with nothing it figures it out.
nice idea, but nowadays we can ask codex oe claude to use each other to perform tasks on a repo or file and the output from one can be used by/for other. hows is this different from that ?
Interesting but it feels like giving an agent the HTTP protocol documentation and a raw socket tool instead of curl, to make requests. Very wasteful in tokens and ultimately computation.
This is of course because we are limited to using agents interactively to avoid paying the API rates.
An agent harness with this integrated would be much more efficient.
I get that there's only so many good names, but Murmur is the name of the Mumble server implementation [https://www.mumble.info/], which is also communication. That makes this name collision worse than others.
I think I have a poor man’s version of this? I run agents in tmux that are aware of each other, and communicate by reading panes using tmux cli tools.
But, I do think something like this is a good solution! Thanks for sharing
Mind sharing your solution a bit more - how do they trigger actions from each other for instance (or is that prompted by you asking them to read from each other)? Asking because OP is interesting but seems a bit overambitious for my own usecases (two pairs of claude + codex agents such that each pair should talk between each other but not across pairs; all running on the same machine).
“You’re in a tmux pane, and the adjacent pane is another agent building a PR for me. Watch it and make sure it doesn’t get stuck before it opens the PR.”
It might make a tmux helper script, a tmux skill might help, but coming in with nothing it figures it out.
nice idea, but nowadays we can ask codex oe claude to use each other to perform tasks on a repo or file and the output from one can be used by/for other. hows is this different from that ?
Interesting but it feels like giving an agent the HTTP protocol documentation and a raw socket tool instead of curl, to make requests. Very wasteful in tokens and ultimately computation.
This is of course because we are limited to using agents interactively to avoid paying the API rates. An agent harness with this integrated would be much more efficient.